RefinedToolCall-V5-3B model card

README.md

@@ -0,0 +1,101 @@
+---
+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
+  - multi-turn
+  - agentic
+  - hermes
+  - reasoning
+  - qwen2
+---
+
+# 🛠️🧠 RefinedToolCall-V5-3B
+
+### A 3B model that *reasons* and *calls tools* — and actually holds a multi-turn conversation.
+
+**Math-grade reasoning · real function calling · multi-turn agentic · 2.5 GB · runs on your laptop.**
+
+> `ollama run refinedneuro/refinedtoolcallv5-3b`
+
+---
+
+## Why it's different
+
+Most 3B tool-callers nail a single function call and then fall apart the moment the task spans
+several turns. **RefinedToolCall-V5** was built specifically to fix that — and the numbers moved on
+*every* axis at once, not just the one we were targeting.
+
+- 🔁 **Multi-turn agentic that actually works** — **~3.7× better** at stateful, multi-step
+  tool-use (Berkeley Function-Calling Leaderboard `multi_turn`) than where we started.
+- 🛠️ **Sharper single-turn calling** — **70.7%** on BFCL single-turn (held-out), our best ever.
+- 💪 **Recovers from tool errors** — **0.896** recovery rate; it diagnoses failures instead of
+  looping on them.
+- 🧮 **Reasoning fully intact** — **AIME-2024 pass@8 0.933**, unchanged by all the tool training.
+- ⚡ **Tiny & local** — 3B params, **2.5 GB** Q6_K, one command on Ollama, no GPU required.
+- 🆓 **Apache-2.0** — use it, ship it, fine-tune it.
+
+---
+
+## The receipts (all held-out, canary-gated)
+
+| capability | this model |
+|---|---|
+| 🔁 Multi-turn agentic (BFCL `multi_turn`, k=3) | **0.220 avg / 0.298 pass@3** |
+| 🛠️ Single-turn function calling (BFCL, held-out) | **0.707** |
+| 🩹 Recovery from tool errors (n=250) | **0.896** |
+| 🧮 Reasoning (AIME-2024 pass@8) | **0.933** |
+
+Every number is the **best across five fine-tuning rounds** — multi-turn, single-turn, recovery,
+*and* reasoning all peaked together.
+
+---
+
+## How we got here (and why it generalizes)
+
+We didn't just throw data at it. Five disciplined rounds, each one gated so it could **never**
+regress reasoning or recovery:
+
+1. **Grounding** — stop inventing shell commands; call the actual functions.
+2. **Plan + finish** — think before calling, and know when the turn is done.
+3. **Scale + long context** — harder tasks, up to 24k tokens.
+4. **On-policy self-improvement (the breakthrough)** — the model learns from its *own* successful
+   multi-turn solutions (expert iteration), which broke past the imitation ceiling **and** sharpened
+   single-turn calling and error-recovery as a bonus.
+
+---
+
+## Quick start
+
+**Ollama**
+```bash
+ollama run refinedneuro/refinedtoolcallv5-3b      # latest = Q6_K, 2.5 GB
+```
+> 💡 Use **Q6_K or higher** for tool-calling — lower quants corrupt the call tokens.
+
+**Format:** ChatML + Hermes tools. Each turn the model emits a `<think>` plan → one or more
+`<tool_call>` blocks → a final reply. Recommended: temp 0.6, top_p 0.95, repeat_penalty 1.1.
+
+---
+
+## Great for
+✅ Local/offline agentic tool-use prototypes ✅ Multi-step function-calling assistants
+✅ Math & STEM reasoning ✅ Learning how small agentic models are actually built.
+
+## Be honest with me (research preview)
+⚠️ It's a **3B research preview**. Multi-turn is **dramatically improved (~3.7×) but not solved** —
+very long, open-ended autonomous loops can still write buggy code or mis-plan. A brilliant,
+tiny building block; not yet a drop-in autonomous engineer.
+
+---
+
+*Built on [WeiboAI/VibeThinker-3B](https://huggingface.co/WeiboAI/VibeThinker-3B) +
+[lambda/hermes-agent-reasoning-traces](https://huggingface.co/datasets/lambda/hermes-agent-reasoning-traces).
+Trained with distribution-matched RFT + on-policy expert iteration, every checkpoint gated against
+reasoning/recovery canaries. Apache-2.0.*