Model card: usage, training details, limitations

README.md

@@ -1,22 +1,131 @@
 ---
-language: [en]
-base_model: Qwen/Qwen3.5-9B
-tags: [sft, reasoning, coding, tool-calling, distillation]
+base_model:
+- Qwen/Qwen3.5-9B
+base_model_relation: finetune
+library_name: transformers
 license: apache-2.0
+license_link: https://huggingface.co/Qwen/Qwen3.5-9B/blob/main/LICENSE
+pipeline_tag: text-generation
+language:
+- en
+tags:
+- qwen3.5
+- reasoning
+- tool-calling
+- distillation
+- sft
+- lora
+datasets:
+- nvidia/Nemotron-SFT-Math-v4
+- nvidia/Nemotron-SFT-Competitive-Programming-v2
+- nvidia/Nemotron-SFT-Agentic-v2
+- nvidia/Nemotron-SFT-Instruction-Following-Chat-v3
+- nvidia/Nemotron-SFT-Science-v2
 ---
 
-# FlashModel — Qwen3.5-9B
+# ⚡ FlashModel-Qwen3.5-9B
 
-Text model fine-tuned with LoRA r=128 / RSLoRA, merged into the base weights.
-Training used 9638 examples and 600 optimizer steps, distilled from open-weight teachers
-whose licenses permit training on outputs: DeepSeek-V4-Pro (math, verified answers), DeepSeek-R1-0528
-(competitive programming), GLM-4.6 / DeepSeek-V3.2 (tool calling), GPT-OSS-120B (instruction following),
-GPT-OSS / Kimi-K2 / DeepSeek-V3.2 (science), via NVIDIA Nemotron SFT datasets (CC BY 4.0 / CC BY-SA 4.0).
-Exact runtime versions and the base revision are recorded in `training_metadata.json`.
+A reasoning / coding / tool-calling fine-tune of [Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B),
+distilled from open-weight frontier teachers (DeepSeek-V4-Pro, DeepSeek-R1-0528, GLM-4.6, DeepSeek-V3.2, GPT-OSS-120B, Kimi-K2).
 
-The training data contains reasoning-budget tags (`off`, `low`, `medium`, `high`,
-`xhigh`, `max`, `adaptive`). Their effect and model quality have not yet been benchmarked.
+**GGUF quants (llama.cpp, Ollama, LM Studio): [flashback2k/FlashModel-Qwen3.5-9B-GGUF](https://huggingface.co/flashback2k/FlashModel-Qwen3.5-9B-GGUF)**
 
-Native MTP export is a separate artifact: the base checkpoint's original MTP head
-is paired with this target model. No DSpark heads were trained or exported.
-Speculative-decoding acceptance and speed must be measured on the target runtime.
+Merged BF16 weights, text-only (`Qwen3_5ForCausalLM`, no vision tower, no MTP head).
+
+## Usage
+
+### Transformers
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model_id = "flashback2k/FlashModel-Qwen3.5-9B"
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+model = AutoModelForCausalLM.from_pretrained(model_id, dtype="auto", device_map="auto")
+
+messages = [{"role": "user", "content": "Write a Python function that checks whether a number is prime."}]
+inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
+output = model.generate(inputs, max_new_tokens=4096, do_sample=True, temperature=0.6, top_p=0.95, top_k=20)
+print(tokenizer.decode(output[0][inputs.shape[-1]:], skip_special_tokens=True))
+```
+
+Qwen3.5 support requires a recent `transformers` release.
+
+### vLLM / SGLang
+
+```bash
+vllm serve flashback2k/FlashModel-Qwen3.5-9B --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder
+```
+
+### Recommended sampling
+
+These are the base model's official recommendations ([Qwen3.5 model card](https://huggingface.co/Qwen/Qwen3.5-9B)); they were not re-tuned for the fine-tune.
+
+| Mode | temperature | top_p | top_k | min_p | presence_penalty |
+|:--|--:|--:|--:|--:|--:|
+| Thinking, general | 1.0 | 0.95 | 20 | 0.0 | 1.5 |
+| Thinking, precise coding | 0.6 | 0.95 | 20 | 0.0 | 0.0 |
+| Non-thinking, general | 0.7 | 0.8 | 20 | 0.0 | 1.5 |
+
+Always pass `--jinja` so the embedded Qwen3.5 chat template (thinking blocks, tool calls) is used.
+
+### Reasoning-budget tag
+
+Training system prompts started with a reasoning-budget tag chosen from the length of the teacher's reasoning:
+
+```
+<|reasoning_budget|>medium<|/reasoning_budget|>
+```
+
+Values: `off`, `low`, `medium`, `high`, `xhigh`, `max`. The tag's effect on output length **has not been measured yet**; treat it as experimental.
+
+### Reasoning-budget tag
+
+Training system prompts started with a reasoning-budget tag chosen from the length of the teacher's reasoning:
+
+```
+<|reasoning_budget|>medium<|/reasoning_budget|>
+```
+
+Values: `off`, `low`, `medium`, `high`, `xhigh`, `max`. The tag's effect on output length **has not been measured yet**; treat it as experimental.
+
+## About the fine-tune
+
+| | |
+|:--|:--|
+| Base | Qwen/Qwen3.5-9B |
+| Method | LoRA r=128 (RSLoRA, α=32) on attention + MLP projections, merged |
+| Data | 9,638 examples / 45M tokens, loss on assistant turns only |
+| Context in training | up to 16,384 tokens (longer examples dropped, never truncated) |
+| Schedule | 1 epoch, 600 steps, lr 5e-5 cosine |
+| Held-out eval loss | 0.5745 (step 100) → 0.5586 (step 600) |
+
+### Training data and teachers
+
+Only open-weight teachers whose licenses allow training on their outputs:
+
+| Share | Domain | Dataset | Teacher |
+|--:|:--|:--|:--|
+| 40% | Math (answers verified against references) | [nvidia/Nemotron-SFT-Math-v4](https://huggingface.co/datasets/nvidia/Nemotron-SFT-Math-v4) | DeepSeek-V4-Pro |
+| 22% | Competitive programming (Python) | [nvidia/Nemotron-SFT-Competitive-Programming-v2](https://huggingface.co/datasets/nvidia/Nemotron-SFT-Competitive-Programming-v2) | DeepSeek-R1-0528 |
+| 16% | Multi-turn tool calling (judge-filtered) | [nvidia/Nemotron-SFT-Agentic-v2](https://huggingface.co/datasets/nvidia/Nemotron-SFT-Agentic-v2) | GLM-4.6 / DeepSeek-V3.2 |
+| 11% | Instruction following | [nvidia/Nemotron-SFT-Instruction-Following-Chat-v3](https://huggingface.co/datasets/nvidia/Nemotron-SFT-Instruction-Following-Chat-v3) | GPT-OSS-120B |
+| 11% | Science reasoning | [nvidia/Nemotron-SFT-Science-v2](https://huggingface.co/datasets/nvidia/Nemotron-SFT-Science-v2) | GPT-OSS / Kimi-K2 / DeepSeek-V3.2 |
+
+Datasets © NVIDIA, CC BY 4.0 (some Math StackExchange-derived samples CC BY-SA 4.0).
+
+The training run's step count, sample count and eval-loss history are in [`training_metadata.json`](training_metadata.json).
+
+## ⚠️ Status and known limitations
+
+- **No benchmarks yet.** Lower held-out loss means the model imitates the teachers more closely; it does not by itself prove
+  it beats the stock Qwen3.5-9B. Comparative evals are planned and will be added here.
+- **English-centric.** All training data was English. On non-English prompts (e.g. Russian) the model often *reasons* in English.
+- Math and code examples longer than 16k tokens were excluded, which skews those domains toward shorter problems.
+- Inherits the base model's limitations and biases.
+
+## Credits
+
+[Qwen team](https://huggingface.co/Qwen) for Qwen3.5 · [NVIDIA](https://huggingface.co/nvidia) for the Nemotron SFT datasets ·
+DeepSeek, Zhipu AI (GLM), OpenAI (GPT-OSS) and Moonshot AI (Kimi) for open-weight teachers ·
+[ggml-org/llama.cpp](https://github.com/ggml-org/llama.cpp).