---
license: apache-2.0
language:
- en
- zh
tags:
- spiking-neural-network
- snn
- ponder-net
- pretrained
- custom-architecture
- neuronspark
library_name: transformers
---

# NeuronSpark V3 — 1.1B Pretrain (in-progress, step 74000)

**Architecture**: SNN (Spiking Neural Network) decoder with PonderNet-style
adaptive K time-step routing. Custom model, not a transformer variant.

- Hidden dim D = 1024 · K_max = 12 · 24 layers
- ~1.92B parameters (model.safetensors ≈ 2.47 GB bf16)
- Tokenizer vocab = 128387 (multilingual)
- Pretrain step **74000 / 380206** (~21% complete)
- Tokens seen: ~5.6B
- Optimizer: AdamW + DeepSpeed ZeRO-2

This is an **in-training snapshot**; not a finished model.

## Load (inference / fine-tune)

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "Brain2nd/NeuronSpark-V3-1.1B-Pretrain",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(
    "Brain2nd/NeuronSpark-V3-1.1B-Pretrain"
)
```

## Resume pretraining

The `deepspeed/` directory contains 8-rank ZeRO-2 sharded optimizer state.
Use with DeepSpeed launcher to continue training from this exact point:

```bash
git lfs install
git clone https://huggingface.co/Brain2nd/NeuronSpark-V3-1.1B-Pretrain ./ckpt
deepspeed --num_gpus=8 train_pretrain.py \
  --deepspeed_config ds_config.json \
  --resume ./ckpt
```

## Files

| | |
|---|---|
| `model.safetensors` | HF-format bf16 weights |
| `config.json` / `generation_config.json` | model config |
| `tokenizer.json` / `tokenizer_config.json` | tokenizer |
| `modeling_neuronspark.py` / `configuration_neuronspark.py` / `__init__.py` | custom architecture (`trust_remote_code=True`) |
| `deepspeed/` | ZeRO-2 optimizer state (8 ranks, fp32 master + Adam moments) |
| `training_state.pth` | step / tokens_seen / epoch metadata |
| `zero_to_fp32.py` | DeepSpeed helper to consolidate sharded state |