What it is. Nitzotz reads a Hebrew message and answers questions you type about it: pick one of several options, give a score on a scale, or say yes or no to a claim. For every answer it gives a probability you can trust, so you know when it is sure and when it is guessing. It does not write text, so it cannot make things up. It runs on a normal laptop, with no internet connection and no cost per question.
What it is for. Deciding what to do with incoming messages: is this a scam, what kind of message is it, which department should get it, how urgent is it. It is not a chatbot and it is not built for long documents (see Limitations).
In numbers. On 298 Hebrew messages it says correctly whether a message is a scam 92.0% of the time. For context: 70% of those messages are not scams, so a model that always says "not a scam" would score 70.5%. The number that matters more is the ranking: it gives real scams a higher probability than legitimate messages 97% of the time (AUC 0.97).
Try it
Python (the laya library, pip install laya):
import laya
agent = laya.load("BrainboxAI/nitzotz")
q = {"scam": {"type": "noul",
"instructions": "ืืืืืขื ืื ืกื ืืืจืื ืื ืืขื ืืืืืฅ ืขื ืงืืฉืืจ, ืืฉืื ืื ืืืกืืจ ืคืจืืื ืืื ืกืืื ืืืืืืืืช.",
"criteria": {"true": "ืื, ืื ื ืืกืืื ืืจืื",
"false": "ืื, ืื ืืืืขื ืืืืืืืืช, ืื ืื ืืฉ ืื ืงืืฉืืจ ืื ืืงืฉืช ืชืฉืืื"}}}
print(agent.predict("ืืืืืื ืฉืื ืืขืืืืช. ืืฉืืจืืจ ืฉืื 12.90 ืืงืืฉืืจ", q)["answers"]["scam"]["noul"])
Output: 0.8628, the probability that the claim ("the message tries to make the reader click a link, pay or hand
over details for no legitimate reason") is true.
The wording of the question matters. This is the exact question the scam test used, and the numbers on this card are for it. In our tries a shorter wording ("the message is a scam attempt") gave clearly worse answers, for example a high scam probability for a plain verification code. If you change the wording, check it on your own messages first.
laya.exe (the standalone binary from ggmlc releases, no Python needed). Download the Q8 file, start the daemon, then send one JSON request per line; each answer comes back on one line:
hf download BrainboxAI/nitzotz nitzotz-q8_0.gguf --local-dir .
laya.exe daemon nitzotz-q8_0.gguf --device vulkan
{"id":"1","state":"ืืื, ืืคืืืฉื ืืืจ ื-10 ืขืืืื ืืชืืงืฃ?","questions":{"scam":{"type":"noul","instructions":"ืืืืืขื ืื ืกื ืืืจืื ืื ืืขื ืืืืืฅ ืขื ืงืืฉืืจ, ืืฉืื ืื ืืืกืืจ ืคืจืืื ืืื ืกืืื ืืืืืืืืช.","criteria":{"true":"ืื, ืื ื ืืกืืื ืืจืื","false":"ืื, ืื ืืืืขื ืืืืืืืืช, ืื ืื ืืฉ ืื ืงืืฉืืจ ืื ืืงืฉืช ืชืฉืืื"}}}}
{"status":"ready","model":"laya"}
{"model": "nitzotz", "family": "nitzotz", "route": "forced nitzotz", "answers": {"scam": {"type": "noul", "action": {"act_probability": 0.908}, "confidence": 0.9536, "noul": 0.0464}}, "usage": {"input_tokens": 66, "output_tokens": 0, "latency_ms": 121.783}, "id": "1"}
noul is the probability that the claim is true. Several questions in one request are answered together in one pass.
Use --device cpu on a machine without a GPU.
Benchmarks
8 frozen test sets, 3,990 questions in total, locked by checksum before any training data existed. All three models answered exactly the same questions. The two reference points are the other open laya models that read Hebrew: RoeiG/laya-hebrew (licence CC-BY-NC-SA-4.0, non-commercial use only) and laya-multilingual (Apache-2.0).
Full table. Accuracy, the best in each row in bold. The last two columns say whether Nitzotz's difference from that model is real or could be luck (a paired exact McNemar test on the same questions): "better" or "worse" means p < 0.05, "tie" means the difference could be chance.
| Test (questions) | Nitzotz | RoeiG (non-commercial) | laya-multilingual | Chance | Nitzotz vs RoeiG | Nitzotz vs laya-multilingual |
|---|---|---|---|---|---|---|
| Scam or not? (298 messages) | 92.0% | 32.2% | 50.3% | 50.0% | better p<0.001 |
better p<0.001 |
| same, hard cases only (65) | 83.1% | 32.3% | 35.4% | 50.0% | better p<0.001 |
better p<0.001 |
| Message type, 6 options (298) | 77.8% | 52.7% | 20.8% | 16.7% | better p<0.001 |
better p<0.001 |
| same, hard cases only (65) | 67.7% | 44.6% | 24.6% | 16.7% | better p=0.01 |
better p<0.001 |
| Support ticket type, 5 options (30) | 90.0% | 80.0% | 56.7% | 20.0% | tie p=0.45 |
better p=0.01 |
| Ticket urgency, 5 levels (30) | 70.0% | 36.7% | 33.3% | 20.0% | better p=0.03 |
better p=0.02 |
| Paying customer? yes/no (30) | 80.0% | 56.7% | 50.0% | 50.0% | better p=0.02 |
tie p=0.06 |
| Voice command intent, 20 options (500) | 90.0% | 72.6% | 47.4% | 5.0% | better p<0.001 |
better p<0.001 |
| Voice command intent, 4 options (500) | 97.2% | 90.8% | 69.4% | 25.0% | better p<0.001 |
better p<0.001 |
| News topic, 7 options (204) | 79.4% | 82.3% | 66.2% | 14.3% | tie p=0.36 |
better p=0.001 |
| Does the passage support this answer? (600) | 94.8% | 94.2% | 48.8% | 50.0% | tie p=0.69 |
better p<0.001 |
| Plausible answer the passage does not give (600) | 89.2% | 53.0% | 53.7% | 50.0% | better p<0.001 |
better p<0.001 |
| Reading comprehension, 4 options (900) | 63.2% | 75.4% | 31.4% | 25.0% | worse p<0.001 |
better p<0.001 |
On Belebele reading comprehension Nitzotz scores 63.2%, below RoeiG/laya-hebrew (75.4%); Nitzotz is built for message decisions, not long-passage comprehension.
How to read it:
- MASSIVE (voice commands): Nitzotz was trained on MASSIVE's training commands. The test commands are different, but written by the same people in the same style, so this test is easier for Nitzotz than for the others.
- HeQ: Nitzotz was trained on HeQ's training passages. The test passages are different ones, and training passages that overlapped a test passage were removed.
- The support-ticket rows have only 30 questions. No difference there is reliable.
- The spam and scam test has a lean: 70% of its messages are not scams. The "chance" line (50%) is a coin flip, not the best blind strategy.
Why you can trust it
The probabilities mean something. Take all the answers where Nitzotz said it was about 70 to 80% sure, and count
how many were right: the chart above does that for every confidence level, over 3,990 test questions.
Where the dots sit above the line, Nitzotz is more often right than it claims (it is modest); where they sit below, it
is over-confident. This is what lets you set thresholds (see the next section). The per-question-type temperatures
were fitted on 4,825 held-out training items, never on the test sets (calibration.json).
It reads the text. With the message removed and only the question left, its accuracy falls to 70.5% on the scam question and 7.7% on message type. So the answers come from the message, not from the wording of the question.
Rewording the question rarely changes the answer. We asked 1,786 test questions in 6 different wordings with the same meaning. On 5.0% of them the answer was not the same in all 6. No wording makes it give one fixed answer to every item of a test. The exception is the urgency question (see Limitations).
It is fast on ordinary hardware. On one laptop (Intel Core Ultra 9 285H laptop, built-in Arc 140T GPU, Windows 11), one question at a time:
| Runtime | Median over 50 check questions (about 135 tokens) | Short message (38 tokens) | Long input (430 tokens) |
|---|---|---|---|
| laya.exe, Q8 file, GPU (Vulkan) | 48 ms | 37 ms | 127 ms |
| laya.exe, F16 file, GPU (Vulkan) | 50 ms | 43 ms | 111 ms |
| Python (laya), GPU (PyTorch XPU) | 60 ms | 44 ms | 139 ms |
| laya.exe, Q8 file, CPU only (16 threads) | 309 ms | 160 ms | 1137 ms |
| Python (laya), CPU only | 224 ms | 112 ms | 1022 ms |
The short and long columns repeat one fixed input 30 times after 5 warm-up calls. Timings on this laptop change a lot from one session to another (an earlier measurement of the same setup was several times slower), so treat these as rough.
The GGUF files give almost the same answers as the Python model. Compared with the full-precision Python model on the CPU:
| File, device | Same top answer, 50 questions | Largest probability gap | Same top answer, 596 spam questions | Largest gap | Average gap |
|---|---|---|---|---|---|
| Q8, GPU | 50/50 | 0.0375 | 596/596 | 0.0166 | 0.00130 |
| F16, GPU | 50/50 | 0.0017 | 596/596 | 0.0015 | 0.00020 |
| Q8, CPU | 50/50 | 0.0549 | 595/596 | 0.0207 | 0.00192 |
| F16, CPU | 50/50 | 0.0032 | 596/596 | 0.0012 | 0.00018 |
A gap of 0.01 means, for example, 0.83 against 0.84. The few questions where the top answer changes are ones where the two best answers were almost tied. Over both spam questions the Q8 file on the GPU is right 84.7% of the time, against 84.7% for the Python model; the F16 file is closer (84.7%).
Use it in your business
The idea ("Ramzor", traffic light). Every incoming message gets one or more quick questions. The probability decides what happens next:
- Green (confident it is fine): handle automatically, file it, tag it, route it.
- Yellow (not sure): send it to a person, or to a large language model if you use one.
- Red (confident it is a scam): block or quarantine it.
The drawing is a worked example on the 298 test messages. The upper threshold, 0.49, is the one chosen for the scam question on 1188 held-out training messages (never on the test); the lower one, 0.35, was picked by hand. With them, 207 messages go to green (13 of them are in fact scams), 9 to yellow (2 scams), and 82 to red (9 of them are in fact legitimate). So red should mean "quarantine and check", not "delete". Pick your own thresholds on a sample of your own messages, and decide how many mistakes in green and red you can live with. At the 0.49 threshold alone, the scam answer is right 92.3% of the time overall and 83.1% on the 65 hard cases (at 0.50: 92.0% and 83.1%).
The model is cheap enough to run on every message. The person (or the LLM) only sees the yellow part. Do not use it as the only line of defence for decisions that can hurt someone.
Training data and transparency
Two training stages, on a rented GPU:
- Learning to read. HalleluBERT-large was first trained to find the answer to a question inside a passage, on 27,085 HeQ training questions (CC BY 4.0). Passages that overlapped a test passage were removed.
- Learning to decide. A laya decision head was put on top and the whole model was trained on 116,854 items (112,029 for training, 4,825 held out to pick the best of 2 passes and to fit the temperatures).
| Source | Items | Licence | How it was made |
|---|---|---|---|
| Synthetic: Message type, 6 classes | 14,000 | teacher output, project-owned (DeepSeek MIT; Gemma Apache-2.0) | written by DeepSeek V4.1 Flash, labelled independently by both models |
| Synthetic: Routing to a department (3 to 6 options) | 8,750 | teacher output, project-owned (DeepSeek MIT; Gemma Apache-2.0) | written by DeepSeek V4.1 Flash, labelled independently by both models |
| Synthetic: Yes/no claims about a message | 8,750 | teacher output, project-owned (DeepSeek MIT; Gemma Apache-2.0) | written by DeepSeek V4.1 Flash, labelled independently by both models |
| Synthetic: Urgency, 5 levels | 3,500 | teacher output, project-owned (DeepSeek MIT; Gemma Apache-2.0) | written by DeepSeek V4.1 Flash, labelled independently by both models |
| HeQ: is the proposed answer supported (yes/no) | 6,000 | CC BY 4.0 | the dataset's gold labels, train split only |
| HeQ: plausible answer to an unanswerable question | 9,750 (4,750 of them repeated copies) | CC BY 4.0 | the dataset's gold labels, train split only |
| HeQ: 4-option reading | 5,000 | CC BY 4.0 | the dataset's gold labels, train split only |
| MASSIVE he-IL: voice command intent | 7,000 | CC BY 4.0 | the dataset's gold labels, train split only |
| Scam or not, disguised scams and real messages that look suspicious | 6,000 | model output, project-owned (DeepSeek MIT; Gemma Apache-2.0) | written by DeepSeek V4.1 Flash from scenario outlines written with GPT; kept only if both labellers agreed with the writer |
| Message type of the same messages | 5,374 | model output, project-owned | the type both labellers agreed on |
| Copies of existing items with the question reworded | 21,199 | as the original item; the wordings were written with GPT | question or options in one of 376 alternative wordings written with GPT; the label comes from the original item |
| Reading, 4 options ("which is NOT", reworded answers, several sentences) | 6,825 | passages: FineWeb-2, ODC-By 1.0; questions: model output, project-owned | written by DeepSeek V4.1 Flash, checked by Gemma 4 31B with the passage and again without it; dropped if it could be answered without the passage |
| Topic of a passage, 7 options | 4,000 | passages: FineWeb-2, ODC-By 1.0; labels: model output | labelled by DeepSeek V4.1 Flash and Gemma 4 31B |
| Is the sender an existing paying customer (yes/no) | 575 | model output, project-owned; question wordings written with GPT | support messages written by DeepSeek V4.1 Flash, checked by Gemma 4 31B |
| Scam or not: warnings about scams, short scams without a link, and legitimate look-alikes | 3,746 | model output, project-owned (DeepSeek MIT; Gemma Apache-2.0); question wordings written with GPT | written by DeepSeek V4.1 Flash; kept only if both labellers agreed with the writer, except that a scam only DeepSeek recognised was kept with a softer label (70%) |
| Message type of the same messages | 1,901 | model output, project-owned | the type both labellers agreed on, only where it fits the scam decision |
| "None of the options": existing training items with one option changed | 4,484 | as the original item (MASSIVE and HeQ: CC BY 4.0; reading: FineWeb-2 passages, ODC-By 1.0) | the right answer removed and a "none of the options" choice added; in a third of them a wrong option was removed instead, so "none" is wrong there |
- Synthetic messages (35,000 items). DeepSeek V4.1 Flash (MIT) wrote Israeli-style SMS, WhatsApp and email messages from a plan (intended label, topic, tone, varied fake phone numbers and links). DeepSeek and Gemma 4 31B (Apache-2.0) then each labelled every message on their own. A message was kept only if both agreed; the two models agreed on 94% of them, and 93.7% of the 37,429 written messages were kept. Both ran through DeepInfra (via OpenRouter), with data retention off and "thinking" mode off. The synthetic data is not published.
- Reading, hard scams and support tickets. 6,825 reading questions on Hebrew web passages from FineWeb-2 (ODC-By 1.0): "which of these is NOT true or NOT mentioned" (each paired with a positive question on the same passage), answers said in other words, and answers that need several sentences. DeepSeek V4.1 Flash wrote them and Gemma 4 31B checked each one twice, with the passage and without it; a question that could be answered without the passage was dropped. 6,000 messages that are hard to tell apart (disguised scams, and real messages that look suspicious), written by DeepSeek and kept only if both labellers agreed with each other and with the writer. 4,000 passages labelled with a topic, 575 support messages for the paying-customer question, and 4,750 repeated HeQ items so that the HeQ skills keep their weight in the mix.
- Scam patterns seen in real messages. Real Hebrew messages showed two weak spots: genuine warnings about scams (from banks, the police, companies) flagged as scams, and short scams without a link missed. So DeepSeek V4.1 Flash wrote 3,746 more messages: 1,000 legitimate warnings about scams, 1,476 short scams without a link (a small unpaid debt or toll, a "friend" with a new number, a gift, a fake payment confirmation, a payment app), 389 pairs of a scam and a warning about that same scam, and 492 legitimate messages that look like those scams. The keep rule changed for these: a message was kept if both labellers agreed with the writer, and a scam that only DeepSeek recognised was also kept, with a softer label (70% instead of close to 100%); 134 messages are of that kind. Messages that resembled one of the real messages we checked were dropped before labelling, and the real messages themselves were never used for training. 134 existing training rows (67 messages: a request from a "new number" or for a small debt, with no link) were relabelled as scams after the labellers called them scams (with the 70% label where only DeepSeek did). 4,484 "none of the options" items were made from existing MASSIVE, HeQ and reading training items: in 2,990 the right answer was removed and a "none of the options" choice added, and in 1,494 a wrong option was removed instead, so "none" is wrong there.
- Written with GPT. Part of the training data was written with OpenAI's GPT, through our ChatGPT subscription: 376 alternative wordings of the questions and 34 sets of alternative answer options, used in 35,189 training items (including all 575 paying-customer items), and 120 short scenario outlines from which DeepSeek V4.1 Flash wrote 6,000 messages. GPT wrote no message and no label. In total 37,849 of the 116,854 training items (32%) use text written with GPT. GPT output is covered by OpenAI's terms of use, not by an open licence.
- Open data. HeQ v1.1 (CC BY 4.0; about half of its questions are on Geektime articles, shared by the HeQ authors under the same licence) and MASSIVE he-IL (CC BY 4.0), training splits only. FineWeb-2 Hebrew (ODC-By 1.0) passages for the reading and topic items.
- No leaks from the tests. Every training item was compared with every test question; anything sharing an 8-word run with a test text was dropped, and so were MASSIVE commands equal to a test command. The reading, topic, hard-scam, paying-customer and reworded items were also checked with stricter rules: every passage against every test text (any shared 6-word run), and every question, option and wording against every test question and every reworded test question (exact match and character similarity).
- Not used: no output of any other closed commercial chatbot, no DICTA model, no non-commercial or share-alike data. GPT, a closed commercial model, was used only as described above.
- Size: encoder 357.1M parameters (HalleluBERT-large, MIT, fine-tuned); decision head 26.5M parameters, trained from scratch.
About the tests.
| Test | Questions | Source and licence |
|---|---|---|
| Spam and business messages | 596 (298 messages, 2 questions each) | in-house, Israeli SMS, WhatsApp and email style, 65 hard cases |
| Support tickets (triage30) | 90 (30 tickets, 3 questions each) | in-house |
| MASSIVE he-IL, 20 and 4 options | 500 + 500 | MASSIVE test split, CC BY 4.0 |
| SIB-200 news topic | 204 | CC BY-SA 4.0, used for testing only |
| Belebele reading | 900 | CC BY-SA 4.0, used for testing only |
| HeQ verify and unanswerable | 600 + 600 | HeQ v1.1 test split, CC BY 4.0 |
| "None of the options" (reported apart, see Limitations) | 400 (200 where "none" is right, 200 where it is wrong) | MASSIVE (CC BY 4.0) and Belebele (CC BY-SA 4.0) test questions with a "none of the options" choice added, used for testing only |
Caveats that change how much to trust the numbers:
- The spam and business test was written by an AI model and checked by an AI model, not by a person. Real inboxes will look different. It was frozen after that review (4 labels changed, 2 messages removed).
- Three training runs, three random seeds; this is one of them. On the held-out training items the three are almost level (95.0% here against 95.2% and 95.2%). This one was picked because it did best on the 188 held-out messages of the added scam data (98.4% against 97.3% and 97.3%) and on the real messages we checked, not on the tests in this card. The three runs are close on the large tests: scam or not 92.0% here against 92.0% and 90.6%, news topic 79.4% against 82.8% and 80.4%, reading 63.2% against 60.9% and 63.1%. This run and each of the other two give the same answer on 90.1% to 90.6% of all test questions. On the 30-ticket tests they differ more: ticket urgency 70.0% here against 73.3% and 76.7%.
- The training labels come from two AI models. Where both are wrong in the same way, Nitzotz learned their mistake.
- HeQ's wrong answers in the test were picked by code, not checked by a person.
Limitations
- Reading comprehension of longer passages is limited. Belebele: 63.2%, where a blind guess gets 25%, and still below the best other laya model on this test (see Benchmarks). When the right answer is written word for word in the passage it gets 77% (252 questions); when the answer is said in other words it gets 58% (648 questions); on "which of these is NOT" questions 59% (158 questions). Do not ask it whether a long document supports a claim.
- Checking an answer against a passage (HeQ) works. 94.8% with the passage; with the passage removed it falls to 51.0%, a coin flip. So on this kind of question it really reads the passage.
- Numbers, dates, amounts and rules: not trained and not measured. Compute them in code and pass the result in.
- Hard cases are still the weak spot: scams written to look legitimate (a "supplier" changing bank details, the "CEO" asking for a transfer) and real messages that look like scams (a real bank alert with a link, a real verification code). On the 65 hard cases the scam question is right 83.1% of the time (always answering "not a scam" there would give 70.8%), with AUC 0.84. At the chosen threshold it still misses 5 of the 19 scams written to look legitimate, and flags 6 of the 46 real messages that look like scams.
- Urgency is subjective, and its answer depends on the wording. Even the two teacher models matched the intended urgency only about 62 to 64% of the time. When the urgency question is asked in other words, the answer changes on 63% of the 30 test tickets.
- The in-house tests and much of the training data were written by AI models, not by people. This includes the spam and business test, the support-ticket test and the FineWeb-2 reading questions. Real messages will look different.
- The support-ticket tests are small: 30 tickets per question, so one ticket moves a score by 3.3 points.
- Sarcasm and irony are probably read literally. Not measured.
- 512 tokens (roughly 300 to 400 Hebrew words) per question. A longer message is cut from the end without a warning.
- Hebrew only. Not trained or tested on English or Arabic.
- Not a safety system on its own. It makes mistakes in both directions. Keep a person in the loop for anything that can hurt someone.
- When a "none of the options" choice is added, it picks it too often. Measured on a separate frozen test of 400 questions, MASSIVE and Belebele test questions rebuilt with a "none of the options" choice. When "none" is the right answer, it picks it 76% of the time. When the right answer is in the list, it still picks "none" on 31% of the questions and gets 61% of them right (77% on voice commands, 45% on reading questions). With the text removed it picks "none" almost every time (98%). If you offer such a choice, test it on your own questions first.
- Real messages: not measured on an independent set yet. Training data was added for genuine warnings about scams and for short scams without a link, the two weak spots real messages showed. A measurement on an independent set of real messages is still pending, so this card gives no number for real messages.
Licence and attribution
Apache-2.0 for the weights, the GGUF files and the code. Commercial use is allowed. Built on:
- HalleluBERT-large: the encoder, MIT.
- laya (NandhaKishorM, Convai Innovations): the decision-head architecture and the runtime, Apache-2.0. No laya weights are used.
- HeQ: CC BY 4.0, by Webiks for MAFAT and the Israeli National NLP Program (NNLP-IL); includes Geektime passages.
- MASSIVE: CC BY 4.0, Amazon (FitzGerald et al., 2022).
- DeepSeek V4.1 Flash (MIT) and Gemma 4 31B (Apache-2.0), used through DeepInfra as data writer and labellers.
- FineWeb-2 (Hebrew): ODC-By 1.0, Hugging Face; the passages of the reading and topic items.
- OpenAI GPT, through a ChatGPT subscription (OpenAI terms of use): question wordings and scenario outlines, as described in the training data section.
- ggmlc for the GGUF files.
- Belebele and SIB-200 (CC BY-SA 4.0) were used only to test, never to train.
The full notice is in NOTICE.
ื ืืฆืืฅ, ืืขืืจืืช
ืื ืื. ื ืืฆืืฅ ืงืืจื ืืืืขื ืืขืืจืืช ืืขืื ื ืขื ืฉืืืืช ืฉืืชื ืืงืืืืื ืขืืื: ืืืืืจ ืืืช ืืืื ืืคืฉืจืืืืช, ืืชืช ืฆืืื ืืกืืื, ืื ืืขื ืืช ืื ืื ืื ืขื ืืขื ื. ืขื ืื ืชืฉืืื ืืื ื ืืชื ืืกืชืืจืืช ืฉืืคืฉืจ ืืกืืื ืขืืื, ืื ืฉืืืืขืื ืืชื ืืื ืืืื ืืืชื ืืื ืื ืืฉ. ืืื ืื ืืืชื ืืงืกื, ืืืื ืืื ืื ืืืื ืืืืฆืื ืืืจืื. ืืื ืจืฅ ืขื ืืืฉื ื ืืื ืจืืื, ืืื ืืื ืืจื ื ืืืื ืชืฉืืื ืขื ืื ืฉืืื.
ืืฉืืื ืื. ืืืืืื ืื ืขืืฉืื ืขื ืืืืขืืช ื ืื ืกืืช: ืืื ืื ืืื ืื, ืืืื ืกืื ืืืืขื ืื, ืืืืื ืืืืงื ืืืขืืืจ, ืืื ืื ืืืืฃ. ืื ืื ืฆ'ืืืืื, ืืืื ืื ืื ืื ืืืกืืืื ืืจืืืื (ืจืื ืืืืืืช).
ืืืกืคืจืื. ืขื 298 ืืืืขืืช ืืขืืจืืช ืืื ืงืืืข ื ืืื ืื ืืืืืขื ืืื ืืื ืื ื-92.0% ืืืืงืจืื. ืืฉืืื ืคืจืืคืืจืฆืื: 70% ืืืืืืขืืช ืืืื ืื ืื ืืื ืื, ืื ืฉืืืื ืฉืชืืื ืขืื ื "ืื ืืื ืื" ืืื ืืงืื 70.5%. ืืืกืคืจ ืฉืืฉืื ืืืชืจ ืืื ืืืืจืื: ืืื ื ืืชื ืืืื ืื ืืืืชืืช ืืกืชืืจืืช ืืืืื ืืืชืจ ืืืฉืจ ืืืืืขื ืชืงืื ื ื-97% ืืืืงืจืื (AUC 0.97).
ืื ืกืืช
ืืคืืืชืื (ืืกืคืจืืื laya, pip install laya):
import laya
agent = laya.load("BrainboxAI/nitzotz")
q = {"scam": {"type": "noul",
"instructions": "ืืืืืขื ืื ืกื ืืืจืื ืื ืืขื ืืืืืฅ ืขื ืงืืฉืืจ, ืืฉืื ืื ืืืกืืจ ืคืจืืื ืืื ืกืืื ืืืืืืืืช.",
"criteria": {"true": "ืื, ืื ื ืืกืืื ืืจืื",
"false": "ืื, ืื ืืืืขื ืืืืืืืืช, ืื ืื ืืฉ ืื ืงืืฉืืจ ืื ืืงืฉืช ืชืฉืืื"}}}
print(agent.predict("ืืืืืื ืฉืื ืืขืืืืช. ืืฉืืจืืจ ืฉืื 12.90 ืืงืืฉืืจ", q)["answers"]["scam"]["noul"])
ืืคืื: 0.8628, ืืืกืชืืจืืช ืฉืืืขื ื ("ืืืืืขื ืื ืกื ืืืจืื ืื ืืขื ืืืืืฅ ืขื ืงืืฉืืจ, ืืฉืื ืื ืืืกืืจ ืคืจืืื ืืื ืกืืื ืืืืืืืืช")
ื ืืื ื.
ืื ืืกืื ืฉื ืืฉืืื ืืฉื ื. ืื ืืืืืง ืืฉืืื ืฉืื ืืฉืชืืฉ ืืืื ืืืื ืืืช, ืืืืกืคืจืื ืืืจืืืก ืืื ืื ืขืืื. ืื ืืกืืื ืืช ืฉืื ื ื ืืกืื ืงืฆืจ ืืืชืจ ("ืืืืืขื ืืื ื ืืกืืื ืืื ืื") ื ืชื ืชืฉืืืืช ืืจืืขืืช ืืืจืื, ืืืฉื ืืกืชืืจืืช ืืืืื ืืืื ืื ืืงืื ืืืืืช ืจืืื. ืื ืืฉื ืื ืืช ืื ืืกืื, ืืืืงืื ืืืชื ืงืืื ืขื ืืืืืขืืช ืฉืืื.
ืืื ืคืืืชืื, ืขื laya.exe (ืชืืื ื ืขืฆืืืืช ื-ggmlc). ืืืจืืืื ืืช ืงืืืฅ Q8, ืืคืขืืืื, ืืฉืืืืื ืืงืฉืช JSON ืืืช ืืื ืฉืืจื. ืื ืชืฉืืื ืืืืจืช ืืฉืืจื ืืืช:
hf download BrainboxAI/nitzotz nitzotz-q8_0.gguf --local-dir .
laya.exe daemon nitzotz-q8_0.gguf --device vulkan
{"id":"1","state":"ืืื, ืืคืืืฉื ืืืจ ื-10 ืขืืืื ืืชืืงืฃ?","questions":{"scam":{"type":"noul","instructions":"ืืืืืขื ืื ืกื ืืืจืื ืื ืืขื ืืืืืฅ ืขื ืงืืฉืืจ, ืืฉืื ืื ืืืกืืจ ืคืจืืื ืืื ืกืืื ืืืืืืืืช.","criteria":{"true":"ืื, ืื ื ืืกืืื ืืจืื","false":"ืื, ืื ืืืืขื ืืืืืืืืช, ืื ืื ืืฉ ืื ืงืืฉืืจ ืื ืืงืฉืช ืชืฉืืื"}}}}
{"status":"ready","model":"laya"}
{"model": "nitzotz", "family": "nitzotz", "route": "forced nitzotz", "answers": {"scam": {"type": "noul", "action": {"act_probability": 0.908}, "confidence": 0.9536, "noul": 0.0464}}, "usage": {"input_tokens": 66, "output_tokens": 0, "latency_ms": 121.783}, "id": "1"}
noul ืืื ืืืกืชืืจืืช ืฉืืืขื ื ื ืืื ื. ืืื ืฉืืืืช ืืืงืฉื ืืืช ื ืขื ืืช ืืื, ืืืขืืจ ืืื. ืขื ืืืฉื ืืื ืืจืืืก ืืกื ืืฉืชืืฉืื ื---device cpu.
ืืืื ืื
8 ืกืืื ืฉื ืืืื, 3,990 ืฉืืืืช ืืกื ืืืื, ืฉื ื ืขืื ืืืืืขืช ืืฆืืข ืืคื ื ืฉื ืืฆืจ ืคืจืื ืืืืื ืืื. ืฉืืืฉืช ืืืืืืื ืขื ื ืขื ืืืชื ืฉืืืืช ืืืืืง. ืฉืชื ื ืงืืืืช ืืืฉืืืื ืื ืืืืื laya ืืคืชืืืื ืืืืจืื ืฉืงืืจืืื ืขืืจืืช: RoeiG/laya-hebrew (ืจืืฉืืื CC-BY-NC-SA-4.0, ืืฉืืืืฉ ืื ืืกืืจื ืืืื) ื-laya-multilingual (Apache-2.0).
ืืืืื ืืืืื. ืืืื ืืชืฉืืืืช ืื ืืื ืืช, ืืืื ืืืืชืจ ืืื ืฉืืจื ืืืืืฉ. ืฉืชื ืืขืืืืืช ืืืืจืื ืืช ืืืืจืืช ืื ืืืืื ืฉื ื ืืฆืืฅ ืืืืืื ืืื ืืืืชื ืื ืฉืืืื ืื ืืื (ืืืื McNemar ืืืืืง ืขื ืืืชื ืฉืืืืช ืืืืืง): "ืืื ืืืชืจ" ืื "ืืืฉ ืืืชืจ" ืคืืจืืฉื p ืงืื ื-0.05, "ืชืืงื" ืคืืจืืฉื ืฉืืืืื ืืืื ืืืืืช ืืงืจื.
| ืืืื (ืืกืคืจ ืฉืืืืช) | ื ืืฆืืฅ | RoeiG (ืื ืืกืืจื) | laya-multilingual | ื ืืืืฉ | ื ืืฆืืฅ ืืื RoeiG | ื ืืฆืืฅ ืืื laya-multilingual |
|---|---|---|---|---|---|---|
| ืืื ืื ืื ืื? (298 ืืืืขืืช) | 92.0% | 32.2% | 50.3% | 50.0% | ืืื ืืืชืจ p<0.001 |
ืืื ืืืชืจ p<0.001 |
| ืืืชื ืืืจ, ืจืง ืืืงืจืื ืืงืฉืื (65) | 83.1% | 32.3% | 35.4% | 50.0% | ืืื ืืืชืจ p<0.001 |
ืืื ืืืชืจ p<0.001 |
| ืกืื ืืืืืขื, 6 ืืคืฉืจืืืืช (298) | 77.8% | 52.7% | 20.8% | 16.7% | ืืื ืืืชืจ p<0.001 |
ืืื ืืืชืจ p<0.001 |
| ืืืชื ืืืจ, ืจืง ืืืงืจืื ืืงืฉืื (65) | 67.7% | 44.6% | 24.6% | 16.7% | ืืื ืืืชืจ p=0.01 |
ืืื ืืืชืจ p<0.001 |
| ืกืื ืคื ืืืช ืชืืืื, 5 ืืคืฉืจืืืืช (30) | 90.0% | 80.0% | 56.7% | 20.0% | ืชืืงื p=0.45 |
ืืื ืืืชืจ p=0.01 |
| ืืืืคืืช ืืคื ืืื, 5 ืจืืืช (30) | 70.0% | 36.7% | 33.3% | 20.0% | ืืื ืืืชืจ p=0.03 |
ืืื ืืืชืจ p=0.02 |
| ืืงืื ืืฉืื? ืื/ืื (30) | 80.0% | 56.7% | 50.0% | 50.0% | ืืื ืืืชืจ p=0.02 |
ืชืืงื p=0.06 |
| ืืืื ืช ืคืงืืื ืงืืืืช, 20 ืืคืฉืจืืืืช (500) | 90.0% | 72.6% | 47.4% | 5.0% | ืืื ืืืชืจ p<0.001 |
ืืื ืืืชืจ p<0.001 |
| ืืืื ืช ืคืงืืื ืงืืืืช, 4 ืืคืฉืจืืืืช (500) | 97.2% | 90.8% | 69.4% | 25.0% | ืืื ืืืชืจ p<0.001 |
ืืื ืืืชืจ p<0.001 |
| ื ืืฉื ืฉื ืืืืขื, 7 ืืคืฉืจืืืืช (204) | 79.4% | 82.3% | 66.2% | 14.3% | ืชืืงื p=0.36 |
ืืื ืืืชืจ p=0.001 |
| ืืื ืืงืืข ืชืืื ืืชืฉืืื? (600) | 94.8% | 94.2% | 48.8% | 50.0% | ืชืืงื p=0.69 |
ืืื ืืืชืจ p<0.001 |
| ืชืฉืืื ืกืืืจื ืฉืืงืืข ืื ื ืืชื (600) | 89.2% | 53.0% | 53.7% | 50.0% | ืืื ืืืชืจ p<0.001 |
ืืื ืืืชืจ p<0.001 |
| ืืื ืช ืื ืงืจื, 4 ืืคืฉืจืืืืช (900) | 63.2% | 75.4% | 31.4% | 25.0% | ืืืฉ ืืืชืจ p<0.001 |
ืืื ืืืชืจ p<0.001 |
ืืืื ืช ืื ืงืจื ืฉื Belebele ื ืืฆืืฅ ืืงืื 63.2%, ืคืืืช ื-RoeiG/laya-hebrew (75.4%). ื ืืฆืืฅ ืื ืื ืืืืืืืช ืขื ืืืืขืืช, ืื ืืืื ื ืฉื ืงืืขืื ืืจืืืื.
ืืื ืืงืจืื ืืช ืื:
- MASSIVE (ืคืงืืืืช ืงืืืืืช): ื ืืฆืืฅ ืืืื ืขื ืคืงืืืืช ืืืืืื ืฉื MASSIVE. ืคืงืืืืช ืืืืื ืืืจืืช, ืืื ื ืืชืื ืืืื ืืืชื ืื ืฉืื ืืืืืชื ืกืื ืื, ืืืื ืืืืื ืืื ืงื ืืืชืจ ืื ืืฆืืฅ ืืืฉืจ ืืืืจืื.
- HeQ: ื ืืฆืืฅ ืืืื ืขื ืงืืขื ืืืืืื ืฉื HeQ. ืงืืขื ืืืืื ืืืจืื, ืืงืืขื ืืืืื ืฉืืคืคื ืืงืืข ืืืื ืืืกืจื.
- ืืฉืืจืืช ืฉื ืคื ืืืช ืืชืืืื ืืฉ ืจืง 30 ืฉืืืืช. ืฉืื ืืืื ืฉื ืื ืืืื.
- ืืืื ืืกืคืื ืืืืื ืืืช ืื ืืืืื: 70% ืืืืืืขืืช ืื ืื ืื ืืื ืื. ืงื ื"ื ืืืืฉ" (50%) ืืื ืืืืช ืืืืข, ืื ืืืกืืจืืืื ืืขืืืืจืช ืืืืื ืืืืชืจ.
ืืื ืืคืฉืจ ืืกืืื ืขืืื
ืืืกืชืืจืืืืช ืืฉ ืืฉืืขืืช. ืงืื ืืช ืื ืืชืฉืืืืช ืฉืืื ื ืืฆืืฅ ืืืจ ืฉืืื ืืืื ืืขืจื ื-70 ืขื 80%, ืืกืคืจื ืืื ืืื ืืื ื ืืื ืืช.
ืืืจืฃ ืขืืฉื ืืช ืื ืืื ืจืืช ืืืืืื, ืขื 3,990 ืฉืืืืช ืืืื. ืืฉืื ืงืืืืช ืืขื ืืงื, ื ืืฆืืฅ ืฆืืืง ืืืชืจ ืืื ืฉืืื ืืืืจ (ืืื
ืฆื ืืข). ืืฉืื ืืชืืช, ืืื ืืืื ืืขืฆืื ืืืชืจ ืืื. ืื ืื ืฉืืืคืฉืจ ืืงืืืข ืกืคืื (ืืคืจืง ืืื). ืืืืื ื ืขืฉื ืขื 4,825 ืคืจืืื ืืืืื
ืฉืืืคืจืื ืืจืืฉ, ืืฃ ืคืขื ืื ืขื ืืืืื (calibration.json).
ืืื ืืืืช ืงืืจื ืืช ืืืงืกื. ืืฉืืืืงืื ืืช ืืืืืขื ืืืฉืืืจืื ืจืง ืืช ืืฉืืื, ืืืืืง ืืืจื ื-70.5% ืืฉืืืช ืืืื ืื ืื-7.7% ืืกืื ืืืืืขื. ืืืืืจ ืืชืฉืืืืช ืืืืช ืืืืืืขื, ืื ืืื ืืกืื ืฉื ืืฉืืื.
ื ืืกืื ืืืจ ืฉื ืืฉืืื ืืืขื ืื ืืฉื ื ืืช ืืชืฉืืื. ืฉืืื ื 1,786 ืฉืืืืช ืืืื ื-6 ื ืืกืืืื ืฉืื ืื ืขื ืืืชื ืืฉืืขืืช. ื-5.0% ืืื ืืชืฉืืื ืื ืืืืชื ืืื ืืื 6 ืื ืืกืืืื. ืืฃ ื ืืกืื ืื ืืืจื ืื ืืชืช ืชืฉืืื ืงืืืขื ืืืช ืืื ืืคืจืืืื ืฉื ืืืื. ืืืืฆื ืื ืืืื ืืื ืฉืืืช ืืืืืคืืช (ืจืื ืืืืืืช).
ืืื ืืืืจ ืขื ืืืืจื ืจืืืื. ืขื ืืืฉื ื ืืื ืขื ืืขืื Intel Core Ultra 9 285H ืืืจืืืก ืืืกื ืืืืื ื Arc 140T, ืืืื ืืืก 11, ืฉืืื ืืืช ืืื ืคืขื:
| ืืื ืืจืืฆืื | ืืฆืืื ืขื 50 ืฉืืืืช ืืืืงื (ืืืืืฆืข 135 ืืืงื ืื) | ืืืืขื ืงืฆืจื (38 ืืืงื ืื) | ืงืื ืืจืื (430 ืืืงื ืื) |
|---|---|---|---|
| laya.exe, ืงืืืฅ Q8, ืืจืืืก ืืกื (Vulkan) | 48 ms | 37 ms | 127 ms |
| laya.exe, ืงืืืฅ F16, ืืจืืืก ืืกื (Vulkan) | 50 ms | 43 ms | 111 ms |
| ืคืืืชืื (laya), ืืจืืืก ืืกื (PyTorch XPU) | 60 ms | 44 ms | 139 ms |
| laya.exe, ืงืืืฅ Q8, ืืขืื ืืืื (16 ืชืืืืืื ืื) | 309 ms | 160 ms | 1137 ms |
| ืคืืืชืื (laya), ืืขืื ืืืื | 224 ms | 112 ms | 1022 ms |
ืืขืืืืืช ืฉื ืืืืืขื ืืงืฆืจื ืืืงืื ืืืจืื ืืืชื ืฉืืื ืจืฆื 30 ืคืขืืื, ืืืจื 5 ืืจืฆืืช ืืืืื. ืืืื ืื ืขื ืืืืฉื ืืื ืืฉืชื ืื ืืจืื ืืื ืืคืขืื ืืืคืขืื (ืืืืื ืงืืืืช ืฉื ืืืชื ืืืืจื ืืฆืื ืืืืืช ืคื ืืื), ืื ืืื ืืกืคืจืื ืืงืืจืื.
ืงืืืฆื ื-GGUF ื ืืชื ืื ืืืขื ืืช ืืืชื ืชืฉืืืืช ืืื ืืืื ืืคืืืชืื. ืืืฉืืืื ืืืืื ืืคืืืชืื ืืืื ืขื ืืืขืื:
| ืงืืืฅ, ืืืฉืืจ | ืืืชื ืชืฉืืื ืืืืืื, 50 ืฉืืืืช | ืืคืจืฉ ืืกืชืืจืืช ืืจืื | ืืืชื ืชืฉืืื ืืืืืื, 596 ืฉืืืืช ืกืคืื | ืืคืจืฉ ืืจืื | ืืคืจืฉ ืืืืฆืข |
|---|---|---|---|---|---|
| Q8, ืืจืืืก ืืกื | 50/50 | 0.0375 | 596/596 | 0.0166 | 0.00130 |
| F16, ืืจืืืก ืืกื | 50/50 | 0.0017 | 596/596 | 0.0015 | 0.00020 |
| Q8, ืืขืื | 50/50 | 0.0549 | 595/596 | 0.0207 | 0.00192 |
| F16, ืืขืื | 50/50 | 0.0032 | 596/596 | 0.0012 | 0.00018 |
ืคืขืจ ืฉื 0.01 ืคืืจืืฉื, ืืืฉื, 0.83 ืืื 0.84. ืืฉืืืืช ืืืขืืืช ืฉืืื ืืชืฉืืื ืืืืืืื ืืฉืชื ื ืื ืืืื ืฉืืื ืฉืชื ืืชืฉืืืืช ืืืืืืช ืืื ืืืขื ืฉืืืช. ืืฉืชื ืฉืืืืช ืืกืคืื ืืื ืงืืืฅ Q8 ืขื ืืจืืืก ืืืกื ืฆืืืง ื-84.7%, ืืื 84.7% ืืืืื ืืคืืืชืื. ืงืืืฅ F16 ืงืจืื ืืืชืจ (84.7%).
ืฉืืืืฉ ืืขืกืง
ืืจืขืืื ("ืจืืืืจ"). ืื ืืืืขื ื ืื ืกืช ืืงืืืช ืฉืืื ืืืืจื ืืืช ืื ืืื. ืืืกืชืืจืืช ืืืืืื ืื ืงืืจื ืืืื:
- ืืจืืง (ืืืื ืฉืื ืืกืืจ): ืืืคืื ืืืืืืื, ืชืืืง, ืชืืื, ื ืืชืื.
- ืฆืืื (ืื ืืืื): ืืืืืงื ืฉื ืืื, ืื ืฉื ืืืื ืฉืคื ืืืื ืื ืืชื ืืฉืชืืฉืื ืื.
- ืืืื (ืืืื ืฉืื ืืื ืื): ืืกืืื ืื ืืกืืจ.
ืืฉืจืืื ืืื ืืืืื ืขื 298 ืืืืขืืช ืืืืื. ืืกืฃ ืืขืืืื, 0.49, ืืื ืืกืฃ ืฉื ืืืจ ืืฉืืืช ืืืื ืื ืขื 1188 ืืืืขืืช ืืืืื ืฉืืืคืจืื ืืจืืฉ (ืืฃ ืคืขื ืื ืขื ืืืืื). ืืกืฃ ืืชืืชืื, 0.35, ื ืืืจ ืืื. ืืืชื 207 ืืืืขืืช ืืืืืืช ืืืจืืง (13 ืืื ืื ืืขืฆื ืืื ืื), 9 ืืฆืืื (2 ืืื ืืืช), ื-82 ืืืืื (9 ืืื ืืขืฆื ืชืงืื ืืช). ืืืืืจ ืืืื ืฆืจืื ืืืืืช "ืืกืืจ ืืืืืงื", ืื "ืืืืงื". ืืืจื ืกืคืื ืืฉืืื ืขื ืืืื ืฉื ืืืืืขืืช ืฉืืื, ืืืืืืื ืืื ืืขืืืืช ืืืจืืง ืืืืืื ืืชื ืืืื ืื ืืงืื. ืืกืฃ 0.49 ืืืื, ืชืฉืืืช ืืืื ืื ื ืืื ื ื-92.3% ืืืืงืจืื ืืกื ืืืื ืื-83.1% ืขื 65 ืืืงืจืื ืืงืฉืื (ืืกืฃ 0.50: 92.0% ื-83.1%).
ืืืืื ืืื ืืกืคืืง ืืื ืืืจืืฅ ืืืชื ืขื ืื ืืืืขื. ืืืื (ืื ืืืื ืืฉืคื) ืจืืื ืจืง ืืช ืืืืง ืืฆืืื. ืื ืชืฉืชืืฉื ืื ืืงื ืืื ื ืืืื ืืืืืืืช ืฉืืืืืืช ืืคืืืข ืืืืฉืื.
ื ืชืื ื ืืืืืื ืืฉืงืืคืืช
ืฉื ื ืฉืืื ืืืืื, ืขื ืืจืืืก ืืกื ืฉืืืจ:
- ืืืืื ืืงืจืื. HalleluBERT-large ืืืื ืงืืื ืืืฆืื ืืช ืืชืฉืืื ืืฉืืื ืืชืื ืงืืข, ืขื 27,085 ืฉืืืืช ืืืืื ืฉื HeQ (CC BY 4.0). ืงืืขืื ืฉืืคืคื ืืงืืข ืืืื ืืืกืจื.
- ืืืืื ืืืืืื. ืืขืืื ืืื ื ืจืืฉ ืืืืืืช ืฉื laya, ืืื ืืืืื ืืืื ืขื 116,854 ืคืจืืืื (112,029 ืืืืืื, ื-4,825 ืืืคืจืื ืืจืืฉ ืืื ืืืืืจ ืืช ืืืื ืืืื 2 ืืขืืจืื ืืืืืื ืืช ืืืืคืจืืืจืืช).
| ืืงืืจ | ืคืจืืืื | ืจืืฉืืื | ืืื ื ืืฆืจ |
|---|---|---|---|
| ืกืื ืชืื: ืกืื ืืืืขื, 6 ืกืืืื | 14,000 | ืืคืื ืฉื ืืืืจื ืฉืืื ืื ื (DeepSeek, MIT; Gemma, Apache-2.0) | ื ืืชื ืืืื DeepSeek V4.1 Flash, ืกืืื ืื ืคืจื ืืืื ืฉื ื ืืืืืืื |
| ืกืื ืชืื: ื ืืชืื ืืืืืงื (3 ืขื 6 ืืคืฉืจืืืืช) | 8,750 | ืืคืื ืฉื ืืืืจื ืฉืืื ืื ื (DeepSeek, MIT; Gemma, Apache-2.0) | ื ืืชื ืืืื DeepSeek V4.1 Flash, ืกืืื ืื ืคืจื ืืืื ืฉื ื ืืืืืืื |
| ืกืื ืชืื: ืืขื ืืช ืื/ืื ืขื ืืืืขื | 8,750 | ืืคืื ืฉื ืืืืจื ืฉืืื ืื ื (DeepSeek, MIT; Gemma, Apache-2.0) | ื ืืชื ืืืื DeepSeek V4.1 Flash, ืกืืื ืื ืคืจื ืืืื ืฉื ื ืืืืืืื |
| ืกืื ืชืื: ืืืืคืืช, 5 ืจืืืช | 3,500 | ืืคืื ืฉื ืืืืจื ืฉืืื ืื ื (DeepSeek, MIT; Gemma, Apache-2.0) | ื ืืชื ืืืื DeepSeek V4.1 Flash, ืกืืื ืื ืคืจื ืืืื ืฉื ื ืืืืืืื |
| HeQ: ืืื ืืชืฉืืื ืืืืฆืขืช ื ืชืืืช (ืื/ืื) | 6,000 | CC BY 4.0 | ืชืืืืืช ืืืื ืฉื ืืืืืจ, ืืคืืฆืื ืืืืืื ืืืื |
| HeQ: ืชืฉืืื ืกืืืจื ืืฉืืื ืฉืืื ืื ืชืฉืืื | 9,750 (ืืชืืื 4,750 ืขืืชืงืื ืืืืจืื) | CC BY 4.0 | ืชืืืืืช ืืืื ืฉื ืืืืืจ, ืืคืืฆืื ืืืืืื ืืืื |
| HeQ: ืงืจืืื, 4 ืืคืฉืจืืืืช | 5,000 | CC BY 4.0 | ืชืืืืืช ืืืื ืฉื ืืืืืจ, ืืคืืฆืื ืืืืืื ืืืื |
| MASSIVE he-IL: ืืืื ืช ืคืงืืื ืงืืืืช | 7,000 | CC BY 4.0 | ืชืืืืืช ืืืื ืฉื ืืืืืจ, ืืคืืฆืื ืืืืืื ืืืื |
| ืืื ืื ืื ืื, ืืื ืืืช ืืืกืืืช ืืืืืขืืช ืืืืชืืืช ืฉื ืจืืืช ืืฉืืืืช | 6,000 | ืคืื ืืืืืื, ืฉืืื ืื ื (DeepSeek, MIT; Gemma, Apache-2.0) | ื ืืชื ืืืื DeepSeek V4.1 Flash ืืคื ืชืจืืืฉืื ืฉื ืืชืื ืขื GPT. ื ืฉืืจ ืจืง ืื ืฉื ื ืืืกืื ืื ืืกืืืื ืขื ืืืืชื |
| ืกืื ืืืืืขื ืฉื ืืืชื ืืืืขืืช | 5,374 | ืคืื ืืืืืื, ืฉืืื ืื ื | ืืกืื ืฉืฉื ื ืืืกืื ืื ืืกืืืื ืขืืื |
| ืขืืชืงืื ืฉื ืคืจืืืื ืงืืืืื ืขื ืืฉืืื ืื ืืกืื ืืืจ | 21,199 | ืืื ืืคืจืื ืืืงืืจื. ืื ืืกืืืื ื ืืชืื ืขื GPT | ืืฉืืื ืื ืืืคืฉืจืืืืช ืืืื ื-376 ื ืืกืืืื ืืืืคืืื ืฉื ืืชืื ืขื GPT. ืืชืืืืช ืืงืืื ืืืคืจืื ืืืงืืจื |
| ืงืจืืื, 4 ืืคืฉืจืืืืช ("ืืืื ืืืืืืช ืื", ืชืฉืืื ืืืืืื ืืืจืืช, ืืื ืืฉืคืืื) | 6,825 | ืงืืขืื: FineWeb-2, ODC-By 1.0. ืฉืืืืช: ืคืื ืืืืืื, ืฉืืื ืื ื | ื ืืชื ืืืื DeepSeek V4.1 Flash ืื ืืืง ืืืื Gemma 4 31B ืขื ืืงืืข ืืฉืื ืืื ืืงืืข. ื ืืจืง ืื ืืคืฉืจ ืืื ืืขื ืืช ืืื ืืงืืข |
| ื ืืฉื ืฉื ืงืืข, 7 ืืคืฉืจืืืืช | 4,000 | ืงืืขืื: FineWeb-2, ODC-By 1.0. ืชืืืืืช: ืคืื ืืืืืื | ืกืืื ืืืื DeepSeek V4.1 Flash ื-Gemma 4 31B |
| ืืื ืืฉืืื ืืงืื ืืฉืื ืงืืื (ืื/ืื) | 575 | ืคืื ืืืืืื, ืฉืืื ืื ื. ื ืืกืืื ืืฉืืื ื ืืชืื ืขื GPT | ืคื ืืืช ืชืืืื ืฉื ืืชืื ืืืื DeepSeek V4.1 Flash ืื ืืืงื ืืืื Gemma 4 31B |
| ืืื ืื ืื ืื: ืืืืจืืช ืืคื ื ืืื ืื, ืืื ืืืช ืงืฆืจืืช ืืื ืงืืฉืืจ, ืืืืืขืืช ืืืืืืืืืช ืฉื ืจืืืช ืืืืื | 3,746 | ืคืื ืืืืืื, ืฉืืื ืื ื (DeepSeek, MIT; Gemma, Apache-2.0). ื ืืกืืื ืืฉืืื ื ืืชืื ืขื GPT | ื ืืชื ืืืื DeepSeek V4.1 Flash. ื ืฉืืจ ืจืง ืื ืฉื ื ืืืกืื ืื ืืกืืืื ืขื ืืืืชื, ืืืฅ ืืืื ืื ืฉืจืง DeepSeek ืืืื, ืฉื ืฉืืจื ืขื ืชืืืืช ืจืื ืืืชืจ (70%) |
| ืกืื ืืืืืขื ืฉื ืืืชื ืืืืขืืช | 1,901 | ืคืื ืืืืืื, ืฉืืื ืื ื | ืืกืื ืฉืฉื ื ืืืกืื ืื ืืกืืืื ืขืืื, ืจืง ืืฉืืื ืืชืืื ืืืืืื ืื ืื ืืื ืื |
| "ืืฃ ืืืช ืืืืคืฉืจืืืืช": ืคืจืืื ืืืืื ืงืืืืื ืขื ืืคืฉืจืืช ืืืช ืฉืฉืื ืชื | 4,484 | ืืื ืืคืจืื ืืืงืืจื (MASSIVE ื-HeQ: CC BY 4.0. ืงืจืืื: ืงืืขืื ื-FineWeb-2, ODC-By 1.0) | ืืชืฉืืื ืื ืืื ื ืืืกืจื ืื ืืกืคื ืืคืฉืจืืช "ืืฃ ืืืช ืืืืคืฉืจืืืืช". ืืฉืืืฉ ืืื ืืืกืจื ืืืงืื ืื ืืคืฉืจืืช ืฉืืืื, ืื ืฉ"ืืฃ ืืืช" ืฉืืืื ืฉื |
- ืืืืขืืช ืกืื ืชืืืืช (35,000 ืคืจืืืื). DeepSeek V4.1 Flash (MIT) ืืชื ืืืืขืืช SMS, ืืืืืกืืค ืืืืื ืืกืื ืื ืืฉืจืืื ืืคื ืชืืื ืืช (ืืชืืืืช ืืืชืืื ื ืช, ื ืืฉื, ืืฉืื, ืืกืคืจื ืืืคืื ืืงืืฉืืจืื ืืืืืคืื ืืืืืื ืื). ืืืจ ืื DeepSeek ื-Gemma 4 31B (Apache-2.0) ืกืืื ื ืื ืืืืขื, ืื ืืื ืืื. ืืืืขื ื ืฉืืจื ืจืง ืื ืฉื ืืื ืืกืืืื. ืื ืืกืืืื ืขื 94% ืืืืืืขืืช, ืื ืฉืืจื 93.7% ืืชืื 37,429 ืืืืืขืืช ืฉื ืืชืื. ืฉื ืืื ืจืฆื ืืจื DeepInfra (ืืจื OpenRouter), ืืื ืฉืืืจืช ื ืชืื ืื ืืืื ืืฆื "ืืฉืืื". ืื ืชืื ืื ืืกืื ืชืืืื ืื ืืคืืจืกืืื.
- ืงืจืืื, ืืื ืืืช ืงืฉืืช ืืคื ืืืช ืชืืืื. 6,825 ืฉืืืืช ืงืจืืื ืขื ืงืืขืื ืืืจืฉืช ืืขืืจืืช ืืชืื FineWeb-2 (ODC-By 1.0): "ืืืื ืืืืืืช ืื ื ืืื ื ืื ืื ืืืืืจืช" (ืื ืืืช ืขื ืฉืืื ืืืืืืช ืฆืืืื ืขื ืืืชื ืงืืข), ืชืฉืืืืช ืฉื ืืืจืืช ืืืืืื ืืืจืืช, ืืชืฉืืืืช ืฉืืืจืฉืืช ืืื ืืฉืคืืื. DeepSeek V4.1 Flash ืืชื ืืืชื, ื-Gemma 4 31B ืืืง ืื ืืืช ืคืขืืืื, ืขื ืืงืืข ืืืื ืืงืืข. ืฉืืื ืฉืืคืฉืจ ืืื ืืขื ืืช ืขืืื ืืื ืืงืืข ื ืืจืงื. 6,000 ืืืืขืืช ืฉืงืฉื ืืืืืื ืืื ืืื (ืืื ืืืช ืืืกืืืช, ืืืืืขืืช ืืืืชืืืช ืฉื ืจืืืช ืืฉืืืืช), ืฉื ืืชืื ืืืื DeepSeek ืื ืฉืืจื ืจืง ืื ืฉื ื ืืืกืื ืื ืืกืืืื ืื ืขื ืื ืืื ืขื ืืืืชื. 4,000 ืงืืขืื ืฉืกืืื ื ืืคื ื ืืฉื, 575 ืคื ืืืช ืชืืืื ืืฉืืืช ืืืงืื ืืืฉืื, ื-4,750 ืคืจืืื HeQ ืืืืจืื, ืืื ืฉืืืืืืืช ืฉื HeQ ืืืฉืืจ ืืฉืงื ืืชืขืจืืืช.
- ืืคืืกื ืืื ืื ืฉื ืจืื ืืืืืขืืช ืืืืชืืืช. ืืืืขืืช ืืืืชืืืช ืืขืืจืืช ืืจืื ืฉืชื ื ืงืืืืช ืืืฉืืช: ืืืืจืืช ืืืืชืืืช ืืคื ื ืืื ืื (ืืื ืงืื, ืืืืฉืืจื, ืืืืจืืช) ืฉืกืืื ื ืืืื ืื, ืืืื ืืืช ืงืฆืจืืช ืืื ืงืืฉืืจ ืฉืคืืกืคืกื. ืืื DeepSeek V4.1 Flash ืืชื ืขืื 3,746 ืืืืขืืช: 1,000 ืืืืจืืช ืืืืืืืืืช ืืคื ื ืืื ืื, 1,476 ืืื ืืืช ืงืฆืจืืช ืืื ืงืืฉืืจ (ืืื ืงืื ืื ืืืจื ืฉืื ืฉืืืื, "ืืืจ" ืขื ืืกืคืจ ืืืฉ, ืืชื ื, ืืืฉืืจ ืชืฉืืื ืืืืืฃ, ืืคืืืงืฆืืืช ืชืฉืืืืื), 389 ืืืืืช ืฉื ืืื ืื ืืืืืจื ืขื ืืืชื ืืื ืื, ื-492 ืืืืขืืช ืืืืืืืืืช ืฉื ืจืืืช ืืื ืืืื ืืืช ืืืื. ืืื ืืฉืืืจื ืืฉืชื ื ืืื: ืืืืขื ื ืฉืืจื ืื ืฉื ื ืืืกืื ืื ืืกืืืื ืขื ืืืืชื, ืืืื ืื ืฉืจืง DeepSeek ืืืื ื ืฉืืจื ืื ืื, ืขื ืชืืืืช ืจืื ืืืชืจ (70% ืืืงืื ืงืจืื ื-100%). 134 ืืืืขืืช ืื ืืืกืื ืืื. ืืืืขืืช ืฉืืื ืืืืช ืืืืืืขืืช ืืืืืชืืืช ืฉืืืงื ื ื ืืจืงื ืืคื ื ืืกืืืื, ืืืืืืขืืช ืืืืืชืืืช ืขืฆืื ืื ืฉืืืฉื ืืืืืื. 134 ืฉืืจืืช ืืืืื ืงืืืืืช (67 ืืืืขืืช: ืืงืฉื ื"ืืกืคืจ ืืืฉ" ืื ืขื ืืื ืงืื, ืืื ืงืืฉืืจ) ืชืืืื ืืืืฉ ืืืื ืื ืืืจื ืฉืืืกืื ืื ืงืืขื ืฉืื ืืื ืื (ืขื ืชืืืืช 70% ืืฉืจืง DeepSeek ืงืืข). 4,484 ืคืจืืื "ืืฃ ืืืช ืืืืคืฉืจืืืืช" ื ืื ื ืืคืจืืื ืืืืื ืงืืืืื ืฉื MASSIVE, HeQ ืืงืจืืื: ื-2,990 ืืื ืืชืฉืืื ืื ืืื ื ืืืกืจื ืื ืืกืคื ืืคืฉืจืืช "ืืฃ ืืืช ืืืืคืฉืจืืืืช", ืื-1,494 ืืืกืจื ืืืงืื ืื ืืคืฉืจืืช ืฉืืืื, ืื ืฉ"ืืฃ ืืืช" ืฉืืืื ืฉื.
- ื ืืชื ืขื GPT. ืืืง ืื ืชืื ื ืืืืืื ื ืืชื ืขื GPT ืฉื OpenAI, ืืจื ืื ืื ChatGPT ืฉืื ื: 376 ื ืืกืืืื ืืืืคืืื ืฉื ืืฉืืืืช ื-34 ืกืืื ืืืืคืืื ืฉื ืืคืฉืจืืืืช ืชืฉืืื, ืฉืืืคืืขืื ื-35,189 ืคืจืืื ืืืืื (ืืืื ืื 575 ืืคืจืืืื ืฉื ืฉืืืช ืืืงืื ืืืฉืื), ื-120 ืชืจืืืฉืื ืงืฆืจืื ืฉืืคืืื DeepSeek V4.1 Flash ืืชื 6,000 ืืืืขืืช. GPT ืื ืืชื ืืฃ ืืืืขื ืืืฃ ืชืืืืช. ืืกื ืืืื 37,849 ืืชืื 116,854 ืคืจืืื ืืืืืื (32%) ืืฉืชืืฉืื ืืืงืกื ืฉื ืืชื ืขื GPT. ืืคืื ืฉื GPT ืืคืืฃ ืืชื ืื ืืฉืืืืฉ ืฉื OpenAI, ืื ืืจืืฉืืื ืคืชืื.
- ื ืชืื ืื ืคืชืืืื. HeQ v1.1 (CC BY 4.0. ืืขืจื ืืฆื ืืืฉืืืืช ืฉืื ืขื ืืชืืืช ืฉื Geektime, ืืืืืจื HeQ ืืฉืชืคืื ืืืชื ืืืืชื ืจืืฉืืื) ื-MASSIVE he-IL (CC BY 4.0), ืจืง ืืคืืฆืืื ืืืืืื. ืงืืขืื ืืขืืจืืช ื-FineWeb-2 (ODC-By 1.0) ืืฉืืืืช ืืงืจืืื ืืื ืืฉื.
- ืืื ืืืืคื ืืืืืื ืื. ืื ืคืจืื ืืืืื ืืืฉืืื ืืื ืฉืืืช ืืืื. ืื ืื ืฉืืืืง ืจืฆืฃ ืฉื 8 ืืืืื ืขื ืืงืกื ืืืื ื ืืจืง, ืืื ืคืงืืืืช MASSIVE ืฉืืืืช ืืคืงืืืช ืืืื. ืคืจืืื ืืงืจืืื, ืื ืืฉื, ืืืื ืืืช ืืงืฉืืช, ืืืงืื ืืืฉืื ืืื ืืกืืืื ืืืืืคืืื ื ืืืงื ืื ืืืืืื ืืืืืจืื ืืืชืจ: ืื ืงืืข ืืื ืื ืืงืกื ืืืื (ืื ืจืฆืฃ ืืฉืืชืฃ ืฉื 6 ืืืืื), ืืื ืฉืืื, ืืคืฉืจืืช ืื ืืกืื ืืื ืื ืฉืืืช ืืืื ืืื ื ืืกืื ืืืืคื ืฉื ืฉืืืช ืืืื (ืืชืืื ืืืืืงืช ืืืืืื ืชืืืื).
- ืื ืื ืฉืืืฉ: ืฉืื ืคืื ืฉื ืฆ'ืืืืื ืืกืืจื ืกืืืจ ืืืจ, ืฉืื ืืืื ืฉื DICTA, ืฉืื ื ืชืื ืื ืื ืืกืืจืืื ืื ืืจืืฉืืื "ืฉืืชืืฃ ืืื". GPT, ืืืื ืืกืืจื ืกืืืจ, ืฉืืืฉ ืจืง ืืืชืืืจ ืืืขืื.
- ืืืื: ืืืงืืื 357.1 ืืืืืื ืคืจืืืจืื (HalleluBERT-large, MIT, ืืืื ืืืืฉ). ืจืืฉ ืืืืืืืช 26.5 ืืืืืื ืคืจืืืจืื, ืืืื ืืืคืก.
ืขื ืืืืื ืื.
| ืืืื | ืฉืืืืช | ืืงืืจ ืืจืืฉืืื |
|---|---|---|
| ืืืืขืืช ืกืคืื ืืขืกืงืื | 596 (298 ืืืืขืืช, 2 ืฉืืืืช ืืื ืืืช) | ืคื ืืื, ืืกืื ืื SMS, ืืืืืกืืค ืืืืื ืืฉืจืืื, 65 ืืงืจืื ืงืฉืื |
| ืคื ืืืช ืชืืืื (triage30) | 90 (30 ืคื ืืืช, 3 ืฉืืืืช ืืื ืืืช) | ืคื ืืื |
| MASSIVE he-IL, 20 ื-4 ืืคืฉืจืืืืช | 500 + 500 | ืคืืฆืื ืืืืื ืฉื MASSIVE, CC BY 4.0 |
| SIB-200, ื ืืฉื ืืืืขื | 204 | CC BY-SA 4.0, ืจืง ืืืืืงื |
| Belebele, ืืื ืช ืื ืงืจื | 900 | CC BY-SA 4.0, ืจืง ืืืืืงื |
| HeQ, ืืืืืช ื"ืืื ืชืฉืืื" | 600 + 600 | ืคืืฆืื ืืืืื ืฉื HeQ v1.1, CC BY 4.0 |
| "ืืฃ ืืืช ืืืืคืฉืจืืืืช" (ืืืืื ืื ืคืจื, ืจืื ืืืืืืช) | 400 (200 ืฉ"ืืฃ ืืืช" ื ืืื ื ืืื, 200 ืฉืืื ืฉืืืื ืืื) | ืฉืืืืช ืืืื ืฉื MASSIVE (CC BY 4.0) ื-Belebele (CC BY-SA 4.0) ืขื ืืคืฉืจืืช "ืืฃ ืืืช ืืืืคืฉืจืืืืช" ืฉื ืืกืคื, ืจืง ืืืืืงื |
ืืกืชืืืืืืืช ืฉืืฉื ืืช ืืื ืืกืืื ืขื ืืืกืคืจืื:
- ืืืื ืืกืคืื ืืืขืกืงืื ื ืืชื ืืืื ืืืื AI ืื ืืืง ืืืื ืืืื AI, ืื ืืืื ืืื. ืชืืืืช ืืืืจ ืืืืชืืืช ืืืจืื ืืืจืช. ืืื ื ื ืขื ืืืจื ืืืืืงื ืืื (4 ืชืืืืืช ืฉืื ื, 2 ืืืืขืืช ืืืกืจื).
- ืฉืืืฉ ืจืืฆืืช ืืืืื, ืขื ืฉืืืฉื ืืจืขืื ืืงืจืืืื, ืืื ืืืช ืืื. ืขื ืคืจืืื ืืืืืื ืฉืืืคืจืื ืืจืืฉ ืฉืืืฉ ืืจืืฆืืช ืืืขื ืฉืืืช (95.0% ืืื ืืื 95.2% ื-95.2%). ืืจืืฆื ืืื ื ืืืจื ืื ืืื ืืืืชื ืืืืื ืืืืชืจ ืขื 188 ืืืืืขืืช ืืืืืืงืืช ืื ืชืื ื ืืืื ืื ืฉื ืืกืคื (98.4% ืืื 97.3% ื-97.3%) ืืขื ืืืืืขืืช ืืืืืชืืืช ืฉืืืงื ื, ืื ืืคื ืืืืื ืื ืืืจืืืก ืืื. ืืืืื ืื ืืืืืืื ืฉืืืฉ ืืจืืฆืืช ืงืจืืืืช: ืืื ืื ืื ืื 92.0% ืืื ืืื 92.0% ื-90.6%, ื ืืฉื ืืืืขื 79.4% ืืื 82.8% ื-80.4%, ืืื ืช ืื ืงืจื 63.2% ืืื 60.9% ื-63.1%. ืืจืืฆื ืืื ืืื ืืืช ืืืฉืชืืื ืืืืจืืช ื ืืชื ืืช ืืืชื ืชืฉืืื ืขื 90.1% ืขื 90.6% ืืื ืฉืืืืช ืืืืื. ืืืืื ืื ืฉื 30 ืคื ืืืช ืืืืืืื ืืืืืื ืืืชืจ: ืืืืคืืช ืืคื ืืื 70.0% ืืื ืืื 73.3% ื-76.7%.
- ืชืืืืืช ืืืืืื ืืืืช ืืฉื ื ืืืืื AI. ืืืคื ืฉืฉื ืืื ืืืขืื ืืืืชื ืืืคื, ื ืืฆืืฅ ืืื ืืช ืืืขืืช ืฉืืื.
- ืืชืฉืืืืช ืืฉืืืืืช ืฉื HeQ ืืืืื ื ืืืจื ืืงืื, ืืื ื ืืืงื ืืืื ืืื.
ืืืืืืช
- ืืื ืช ืื ืงืจื ืฉื ืงืืขืื ืืจืืืื ืืืืืืช. Belebele: 63.2%, ืืฉื ืืืืฉ ืขืืืืจ ืืงืื 25%, ืืขืืืื ืคืืืช ืืืืื laya ืืืืจ ืืืื ืืืืชืจ ืืืืื ืืื (ืจืื ืืืื ืื). ืืฉืืชืฉืืื ืื ืืื ื ืืชืืื ืืงืืข ืืืื ืืืืื ืืื ืืงืื 77% (252 ืฉืืืืช). ืืฉืืชืฉืืื ื ืืืจืช ืืืืืื ืืืจืืช ืืื ืืงืื 58% (648 ืฉืืืืช). ืืฉืืืืช "ืืืื ืืืืืืช ืื" 59% (158 ืฉืืืืช). ืื ืชืฉืืื ืืืชื ืื ืืกืื ืืจืื ืชืืื ืืืขื ื.
- ืืืืงื ืื ืงืืข ืชืืื ืืชืฉืืื (HeQ) ืขืืืืช. 94.8% ืขื ืืงืืข. ืืฉืืืืงืื ืืช ืืงืืข ืื ืืืจื ื-51.0%, ืืืืช ืืืืข. ืืืืืจ ืืฉืืืืช ืืืกืื ืืื ืืื ืืืืช ืงืืจื ืืช ืืงืืข.
- ืืกืคืจืื, ืชืืจืืืื, ืกืืืืื ืืืืืื: ืื ืืืื ืขืืืื ืืื ื ืืื. ืืฉืื ืืืชื ืืงืื ืืืขืืืจื ืืช ืืชืืฆืื.
- ืืืงืจืื ืืงืฉืื ืขืืืื ื ืงืืืช ืืชืืจืคื: ืืื ืืืช ืฉื ืืชืื ืืื ืืืืจืืืช ืืืืืืืืืช ("ืกืคืง" ืฉืืืืืฃ ืคืจืื ืืฉืืื ืื ืง, "ืืื ื"ื" ืฉืืืงืฉ ืืขืืจื), ืืืืืขืืช ืืืืชืืืช ืฉื ืจืืืช ืืื ืืื ืื (ืืชืจืื ืืืืชืืช ืืืื ืง ืขื ืงืืฉืืจ, ืงืื ืืืืืช ืืืืชื). ืขื 65 ืืืงืจืื ืืงืฉืื ืฉืืืช ืืืื ืื ืฆืืืงืช ื-83.1% ืืืืงืจืื (ืชืฉืืื ืงืืืขื "ืื ืืื ืื" ืืืืชื ืืงืืืช ืฉื 70.8%), ืขื AUC 0.84. ืืกืฃ ืฉื ืืืจ ืืื ืขืืืื ืืคืกืคืก 5 ืืชืื 19 ืืื ืืืช ืฉื ืืชืื ืืื ืืืืจืืืช ืืืืืืืืืช, ืืืชืจืืข ืขื 6 ืืชืื 46 ืืืืขืืช ืืืืชืืืช ืฉื ืจืืืช ืืื ืืื ืื.
- ืืืืคืืช ืืื ืขื ืืื ืกืืืืืงืืืื, ืืืชืฉืืื ืขืืื ืชืืืื ืื ืืกืื. ืืคืืื ืฉื ื ืืืืื ืืืืจื ืืกืืืื ืขื ืืืืืคืืช ืืืชืืื ื ืช ืจืง ืื-62 ืขื 64% ืืืืงืจืื. ืืฉืฉืืืืื ืืช ืฉืืืช ืืืืืคืืช ืืืืืื ืืืจืืช, ืืชืฉืืื ืืฉืชื ื ื-63% ืืชืื 30 ืคื ืืืช ืืืืื.
- ืืืืื ืื ืืคื ืืืืื ืืืืง ืืืื ืื ืชืื ื ืืืืืื ื ืืชืื ืืืื ืืืืื AI, ืื ืืืื ืื ืฉืื. ืื ืืืื ืืช ืืืื ืืกืคืื ืืืขืกืงืื, ืืช ืืืื ืคื ืืืช ืืชืืืื ืืืช ืฉืืืืช ืืงืจืืื ืขื FineWeb-2. ืืืืขืืช ืืืืชืืืช ืืืจืื ืืืจืช.
- ืืืื ื ืคื ืืืช ืืชืืืื ืงืื ืื: 30 ืคื ืืืช ืืื ืฉืืื, ืื ืฉืคื ืืื ืืืช ืืืืื ืฆืืื ื-3.3 ื ืงืืืืช.
- ืฆืื ืืืช ืืืืจืื ืื ืื ืจืื ื ืงืจืืืช ืืืืืืืช. ืื ื ืืื.
- 512 ืืืงื ืื (ืืขืจื 300 ืขื 400 ืืืืื ืืขืืจืืช) ืืฉืืื. ืืืืขื ืืจืืื ืืืชืจ ื ืืชืืช ืืืกืืฃ ืืื ืืืืจื.
- ืขืืจืืช ืืืื. ืื ืืืื ืืื ื ืืืง ืขื ืื ืืืืช ืื ืขืจืืืช.
- ืื ืื ืืขืจืืช ืืื ื ืืื. ืืื ืืืขื ืืฉื ื ืืืืืื ืื. ืืฉืืืจื ืืื ืืชืืืื ืืื ืืืจ ืฉืืืื ืืคืืืข ืืืืฉืื.
- ืืฉืืืกืืคืื ืืคืฉืจืืช "ืืฃ ืืืช ืืืืคืฉืจืืืืช", ืืื ืืืืจ ืื ืืืชืจ ืืื. ื ืืื ืขื ืืืื ืงืคืื ื ืคืจื ืฉื 400 ืฉืืืืช, ืฉืืืืช ืืืื ืฉื MASSIVE ื-Belebele ืฉื ืื ื ืืืืฉ ืขื ืืคืฉืจืืช "ืืฃ ืืืช ืืืืคืฉืจืืืืช". ืืฉ"ืืฃ ืืืช" ืืื ืืชืฉืืื ืื ืืื ื, ืืื ืืืืจ ืื ื-76% ืืืืงืจืื. ืืฉืืชืฉืืื ืื ืืื ื ื ืืฆืืช ืืจืฉืืื, ืืื ืขืืืื ืืืืจ "ืืฃ ืืืช" ื-31% ืืืฉืืืืช, ืืฆืืืง ืจืง ื-61% ืืื (77% ืืคืงืืืืช ืงืืืืืช, 45% ืืฉืืืืช ืงืจืืื). ืืฉืืืืงืื ืืช ืืืงืกื ืืื ืืืืจ "ืืฃ ืืืช" ืืืขื ืชืืื (98%). ืื ืืชื ืืฆืืขืื ืืคืฉืจืืช ืืื, ืืืงื ืืืชื ืงืืื ืขื ืืฉืืืืช ืฉืืื.
- ืืืืขืืช ืืืืชืืืช: ืขืื ืื ื ืืื ืขื ืกื ืืืชื ืชืืื. ื ืืกืคื ื ืชืื ื ืืืืื ืืืืืจืืช ืืืืชืืืช ืืคื ื ืืื ืื ืืืืื ืืืช ืงืฆืจืืช ืืื ืงืืฉืืจ, ืฉืชื ืื ืงืืืืช ืืืืฉืืช ืฉืืืืขืืช ืืืืชืืืช ืืจืื. ืืืืื ืขื ืกื ืืืชื ืชืืื ืฉื ืืืืขืืช ืืืืชืืืช ืขืื ืื ื ืขืฉืชื, ืืืื ืืืจืืืก ืืื ืืื ืืกืคืจ ืขื ืืืืขืืช ืืืืชืืืช.
ืจืืฉืืื ืืงืจืืืืื
Apache-2.0 ืืืฉืงืืืืช, ืืงืืืฆื ื-GGUF ืืืงืื. ืืืชืจ ืืฉืืืืฉ ืืกืืจื. ืื ืื ืขื:
- HalleluBERT-large: ืืืงืืื, MIT.
- laya (NandhaKishorM, Convai Innovations): ืืืจืืืืงืืืจื ืฉื ืจืืฉ ืืืืืืืช ืืกืืืืช ืืืจืฆื, Apache-2.0. ืื ื ืขืฉื ืฉืืืืฉ ืืืฉืงืืืืช ืฉื laya.
- HeQ: CC BY 4.0, ืฉื Webiks ืขืืืจ ืืคื"ืช ืืชืืื ืืช ื-NLP ืืืืืืืช (NNLP-IL). ืืืื ืงืืขืื ื-Geektime.
- MASSIVE: CC BY 4.0, ืืืืื (FitzGerald ืืืืจืื, 2022).
- DeepSeek V4.1 Flash (MIT) ื-Gemma 4 31B (Apache-2.0), ืืจื DeepInfra, ืืืืชื ืื ืชืื ืื ืืืืกืื ืื.
- FineWeb-2 (ืขืืจืืช): ODC-By 1.0, ืฉื Hugging Face. ืืงืืขืื ืฉื ืฉืืืืช ืืงืจืืื ืืื ืืฉื.
- GPT ืฉื OpenAI, ืืจื ืื ืื ChatGPT (ืชื ืื ืืฉืืืืฉ ืฉื OpenAI): ื ืืกืืื ืฉืืืืช ืืชืจืืืฉืื, ืืืชืืืจ ืืคืจืง ื ืชืื ื ืืืืืื.
- ggmlc ืืงืืืฆื ื-GGUF.
- Belebele ื-SIB-200 (CC BY-SA 4.0) ืฉืืืฉื ืจืง ืืืืืงื, ืืฃ ืคืขื ืื ืืืืืื.
ืืืืืขื ืืืืื ืืงืืืฅ NOTICE.
- Downloads last month
- 144
Model tree for BrainboxAI/nitzotz
Base model
HalleluBERT/HalleluBERT_large










