---
license: mit
base_model: WeiboAI/VibeThinker-3B
tags:
  - code
  - agent
  - tool-use
  - gguf
  - qwen2
library_name: transformers
pipeline_tag: text-generation
---

# SteraVibeThinker

A full fine-tune of [WeiboAI/VibeThinker-3B](https://huggingface.co/WeiboAI/VibeThinker-3B)
(a 3B reasoning model built on the Qwen2.5-3B / Qwen2.5-Coder-3B architecture) on
the ~30k-example **Tiny-Giant** agentic tool-use dataset.

The goal: keep VibeThinker's strong verifiable-reasoning core while teaching it the
deterministic, Hermes/ChatML-style `<tool_call>` agent format used by the
Tiny-Giant harness.

## Files

| File | Description |
|---|---|
| `SteraVibeThinker-Q4_K_M.gguf` | Q4_K_M quantization (~1.8 GB) — for `llama.cpp` / Ollama / LM Studio |
| `SteraVibeThinker-f16.gguf` | f16 GGUF (~5.8 GB) — re-quantize to any level without retraining |
| `raw_weights/` | Full bf16 safetensors HF checkpoint |
| `val_meta.jsonl` | Held-out validation set shipped with the model |

## Training

- **Base:** `WeiboAI/VibeThinker-3B` (MIT, Qwen2.5-3B architecture, ChatML-native)
- **Method:** full fine-tune (not LoRA), bf16 + gradient checkpointing
- **Data:** ~30k Tiny-Giant agentic tool-use conversations
- **Epochs:** 2 · **LR:** 7e-6 (cosine, 3% warmup) · **Seq len:** 4096
- **Loss:** full-sequence (tool results modeled as in-distribution context)

## Prompt format

This model was trained with an **explicit ChatML / Hermes renderer**, not
`tokenizer.apply_chat_template`. Pin the ChatML template explicitly when serving —
do not rely on auto-detection. Tool calls use:

```
<tool_call>
{"name": "<function-name>", "arguments": {...}}
</tool_call>
```

## Inference (llama.cpp)

```bash
llama-cli -m SteraVibeThinker-Q4_K_M.gguf --chat-template chatml
```

## License

MIT, inherited from the VibeThinker-3B base model.