VibeThinker-3B-Hermes — v0.4 (Research Preview)

👉 There's a newer, better model: RefinedToolCall-V5-3B

An on-policy self-RFT successor that improves multi-turn agentic ~3.7× plus single-turn calling (0.707) and error-recovery (0.896) — with reasoning intact. Try it: ollama run refinedneuro/refinedtoolcallv5-3b

⚠️ Research preview / experimental. Read Limitations before use.

A LoRA fine-tune of WeiboAI/VibeThinker-3B (Qwen2-based) that adds Hermes-style function calling (<think>…</think> + <tool_call>…</tool_call>) and — new in v0.4 — meaningfully better multi-turn agentic tool use, while preserving its math/STEM reasoning.

What's new in v0.4

v0.4 targets the multi-turn agentic ceiling that v0.3 could not (v0.3 lifted single-turn function calling +13.8 pts but still failed live multi-turn loops). Using distribution-matched behavior-cloning on gold trajectories from the Berkeley Function-Calling Leaderboard (BFCL) multi_turn tasks — executed against the real stateful backends for genuine tool responses, with a planning <think> and an explicit turn-termination step — multi-turn success improved ~2.5–3× with no regression to reasoning, recovery, or single-turn calling:

metric	v0.3	v0.4	Δ
BFCL multi_turn (held-out multi_turn_base, n=47, k=3)	~0.06	0.156 avg / 0.213 pass@3	▲ ~2.5–3×
BFCL single-turn FC (held-out, n=167)	0.647	0.659	▲
Recovery from tool errors (held-out glm, n=250)	0.888	0.884	≈
AIME-2024 pass@4	0.933	0.933	=
AIME-2024 avg@4	0.783	0.838	▲

All gains measured on held-out splits never seen in training; reasoning/recovery/single-turn were held as canaries with a pre-committed promotion gate (auto-halt on any regression).

How it was made (recipe)

1. Measure on the target distribution. BFCL multi_turn with the official multi_turn_checker (compares final backend state + responses + invoke order) — far stronger than single-turn AST matching. v0.3 baseline here was only 6.4%.
2. Diagnose the real failure. v0.3 reverted to bash habits (invented ls -la instead of calling the provided cd/touch/...), never terminated a turn, and didn't plan call sequences.
3. Gold-trajectory behavior cloning. Execute each task's gold call sequence against the live backend for real tool responses; format as Hermes multi-turn conversations with a planning <think> and a terminating reply (teaching the model to stop and yield).
4. Scale data + context. Add long_context and miss_param categories at up to 24k tokens. This data+context scale-up was the single biggest lever (0.128 → 0.234).
5. Protect the canaries. Mix in single-turn FC + multi-turn recovery + reasoning-replay anchors; gate every checkpoint against FC 0.647 / recovery 0.888 / AIME 0.933 floors.
6. Continue the v0.3 adapter on the singly-quantized base (never QLoRA on a merged model).

Usage

ChatML with Hermes tools. Recommended: temperature 0.6, top_p 0.95, repeat_penalty 1.1, stop on <|im_end|>. The model emits a <think> plan, then one or more <tool_call> blocks, then a final natural-language reply per turn.

Quantized versions

• GGUF / Ollama: use Q6_K or higher for tool-calling (lower quants corrupt call tokens — a measured quantization artifact).

Limitations

Still a 3B research preview. Multi-turn success is improved ~2.5–3× but far from solved (~0.16 avg / 0.21 pass@3 on BFCL multi_turn_base): long, open-ended agent loops remain brittle (occasional buggy generated code, imperfect long-horizon planning). Strong at single-turn function calling and math reasoning; not production-ready for fully autonomous multi-step agentic work.

License

Apache-2.0 — built on VibeThinker-3B and lambda/hermes-agent-reasoning-traces (both Apache-2.0).