🧭 At a glance
| Field | Details |
|---|---|
| 🧠 Base family | Qwen3-4B-Instruct |
| 🪜 Curriculum stage | Stage 3 — Boundary-Guided Online Reinforcement Learning |
| 🧱 Starting point | ToolWeave Stage 2 update 25 |
| 🎛️ Training signal | Verified online data synthesis + multi-turn Progress Reward |
| ✅ Release status | Final ToolWeave Stage 3 model |
ToolWeave Stage 3 expands multi-turn tool-use learning through capability-boundary detection, verified online data synthesis, strict execution and semantic validation, dynamic replay, and combined global/local tool-call credit.
📊 Stage 3 evaluation
The final ToolWeave Stage 3 checkpoint was evaluated on the canonical balanced 400-row held-in set: 100 entries each from Base, Missing Function, Missing Parameter, and Long Context. These values are complete-entry BFCL Multi-Turn accuracies, not the training-time Progress Reward (R_P).
| Model | Overall (%) | Base (%) | Missing Function (%) | Missing Parameter (%) | Long Context (%) | Correct entries |
|---|---|---|---|---|---|---|
| ToolWeave Stage 3 | 55.00 | 62.00 | 59.00 | 46.00 | 53.00 | 220 / 400 |
Because the four categories are balanced, the overall score is their unweighted mean and the complete-entry accuracy over all 400 entries:
(62.00 + 59.00 + 46.00 + 53.00) / 4 = 55.00
🚀 Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "muradil211/ToolWeave_stage3"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")
Tool-use inference requires the model's function schemas and the Qwen3-compatible tool-call format.