🧱 Curriculum role: establish reliable tool-call formatting and environment interaction before progress-reward learning.
🧭 At a glance
| Field | Details |
|---|---|
| 🧠 Base family | Qwen3-4B-Instruct |
| 🪜 Curriculum stage | Stage 1 — Tool-Use Cold Start |
| 📍 Checkpoint | Selected Stage 1 update 25 |
| 🎛️ Training signal | Format and executable tool-use feedback |
| ✅ Release status | Selected checkpoint; not the final ToolWeave model |
ToolWeave Stage 1 is trained to improve parser-compatible tool-call formatting, executable interaction, and stable multi-turn tool-use behavior in the BFCL environment.
📊 Evaluation (eval_400)
This is an internal ToolWeave validation on val_400_combined: 400 examples, with 100 examples each from Base, Long Context, Missing Function, and Missing Parameter. Validation used deterministic decoding (n=1, do_sample=false). The validation split is not included in this model repository, and these results are not official BFCL leaderboard results.
For Stage 1, score is the format-gate score reported by the project evaluation; progress is the mean interaction progress.
| Split | Samples | Score | Format reward | Tool-call reward | Tool-call rate | Progress |
|---|---|---|---|---|---|---|
| Overall | 400 | 1.7007 | 0.8354 | 0.8718 | 0.8875 | 0.3983 |
| Base | 100 | 1.7601 | 0.8633 | 0.9047 | 0.9200 | 0.4368 |
| Long Context | 100 | 1.7292 | 0.8511 | 0.8798 | 0.9000 | 0.4177 |
| Missing Function | 100 | 1.7584 | 0.8627 | 0.9057 | 0.9200 | 0.3889 |
| Missing Parameter | 100 | 1.5550 | 0.7646 | 0.7972 | 0.8100 | 0.3499 |
🚀 Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "muradil211/ToolWeave_stage1"
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.