83 lines
3.0 KiB
Markdown
83 lines
3.0 KiB
Markdown
---
|
||
license: apache-2.0
|
||
base_model: WeiboAI/VibeThinker-3B
|
||
datasets:
|
||
- lambda/hermes-agent-reasoning-traces
|
||
language:
|
||
- en
|
||
pipeline_tag: text-generation
|
||
tags:
|
||
- tool-calling
|
||
- function-calling
|
||
- hermes
|
||
- reasoning
|
||
- gguf
|
||
- llama-cpp
|
||
- research-preview
|
||
---
|
||
|
||
> 👉 **Newer model:** [RefinedToolCall-V5-3B](https://huggingface.co/RefinedNeuro/RefinedToolCallV5-3b) — multi-turn agentic ~3.7×, single-turn 0.707, recovery 0.896, reasoning intact. `ollama run refinedneuro/refinedtoolcallv5-3b`
|
||
|
||
# VibeThinker-3B-Hermes — GGUF (Research Preview)
|
||
|
||
⚠️ **Research preview / experimental.** Read **Limitations** before use. Known issues:
|
||
repetition loops on out-of-distribution multi-turn input, and over-eagerness to call tools.
|
||
|
||
GGUF quantizations of a LoRA fine-tune of
|
||
[`WeiboAI/VibeThinker-3B`](https://huggingface.co/WeiboAI/VibeThinker-3B) that adds
|
||
Hermes-style function calling (`<think>…</think>` + `<tool_call>…</tool_call>`) while
|
||
preserving reasoning. See the full (transformers) model card for training details and
|
||
benchmarks.
|
||
|
||
## Files
|
||
|
||
| File | Quant | Size |
|
||
|---|---|---|
|
||
| `…-Q6_K.gguf` | **Q6_K** | 2.4 GB | min recommended |
|
||
| `…-Q8_0.gguf` | Q8_0 | 3.1 GB |
|
||
| `…-f16.gguf` | F16 | 5.8 GB | best |
|
||
|
||
> **Why only Q6_K and up?** We measured tool-call fidelity on a multi-step agentic task:
|
||
> **Q6_K, Q8_0 and F16 pass; Q3/Q4/Q5 fail** (they emit malformed/incomplete tool calls and
|
||
> can loop). Since this is a tool-calling model, the lower quants were removed to avoid
|
||
> shipping a broken experience. Use **Q6_K** for the best size/quality balance.
|
||
|
||
## Usage (llama.cpp)
|
||
|
||
```bash
|
||
# use Q6_K+ for tool-calling
|
||
llama-cli -m VibeThinker-3B-Hermes-Q6_K.gguf \
|
||
--temp 0.6 --top-p 0.95 --repeat-penalty 1.1 \
|
||
-p "<your Hermes-formatted prompt>"
|
||
```
|
||
|
||
## Recommended settings
|
||
|
||
- **Use a stop on `</tool_call>`** for single-call settings — essential to avoid the model
|
||
appending duplicate/hallucinated calls.
|
||
- temperature 0.6, top_p 0.95, repeat-penalty ≈ 1.1.
|
||
- Stop tokens: `<|im_end|>` (151645) and `<|endoftext|>` (151643).
|
||
- Hermes system prompt with a `<tools>` block; the model emits `<think>…</think>` then
|
||
`<tool_call>\n{"name": …, "arguments": …}\n</tool_call>`.
|
||
|
||
## Benchmarks (summary)
|
||
|
||
- **Reasoning (AIME 2024):** base avg@4 0.842 → this model 0.783; pass@4 0.867 unchanged.
|
||
- **Tool-calling (BFCL single-turn, with stop fix):** live_simple ~60 %, live_multiple ~31 %,
|
||
live_relevance 87.5 %, live_irrelevance ~34 % (over-eager).
|
||
|
||
## Limitations
|
||
|
||
- Repetition loops on OOD multi-turn input (emits the same call many times; avg ~25 calls/turn
|
||
vs ~1.3 expected). Mitigate with `stop=["</tool_call>"]` + repeat-penalty.
|
||
- Over-eager to call tools (does not always decline when it should).
|
||
- Single-turn exact-match is poor without the stop fix.
|
||
- ~6-pt avg@4 reasoning dip vs base; trained 1 epoch on one dataset config.
|
||
|
||
Not recommended for production agents as-is.
|
||
|
||
## License
|
||
|
||
Apache-2.0. Built on `WeiboAI/VibeThinker-3B` and `lambda/hermes-agent-reasoning-traces`
|
||
(both Apache-2.0).
|