---
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
base_model: Qwen/Qwen3-4B
base_model_relation: finetune
tags:
- trajectory2gym
- tool-use
- supervised-fine-tuning
datasets:
- mis2/Trajectory2gym
---

# Trajectory2Gym-4B

Full-parameter supervised post-training for decision-grounded tool use. This checkpoint is the paper's corresponding scale, not a support-depth or component-ablation model.

## Reported evaluation

| Tau1 | Tau2 | BFCL three-group Overall |
|---:|---:|---:|
| 44.22 | 69.17 | 74.85 |

These are recorded paper results, not a fresh benchmark rerun during publication. Tau scores are domain macro averages. Tau1 Airline uses the disclosed seen-task protocol. Tau2 uses the official train/test split, and BFCL Overall equally averages Non-live, Live and Multi-turn groups.

## Use

Use the included tokenizer and native tool schema format with non-thinking generation. For Qwen3 templates supporting it, pass enable_thinking=False to apply_chat_template. The 30B-A3B Instruct-2507 model retains its own native template; do not inject an 8B-specific prefix.

~~~python
from transformers import AutoTokenizer, AutoModelForCausalLM
repo = 'mis2/Trajectory2Gym-4B'
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype='auto', device_map='auto')
~~~

Loading requires hardware memory appropriate to the scale and stored dtype. This release preserves the archived safetensor bytes: float32. No quantization, retraining or numeric conversion was performed for release.

The public candidate corpus is not the exact rendered training view. The 30B-A3B model uses an expanded corpus. Model behavior and benchmark scores depend on the native interfaces, simulator and evaluation settings; this card does not claim cross-protocol equivalence.

## Integrity and license

Weight files total 17,645,742,616 bytes. The source training-view SHA-256 is ab2e580badcec821f5cd3a8937d41366b75220ea6edcf2e9835d1a3e5a0de8cf. RELEASE_MANIFEST.json provides per-file checksums and archive-byte verification details. Internal source paths, credentials, optimizer states, training arguments and raw logs are excluded. Only non-inference path metadata may be removed from JSON configurations, as recorded in the manifest.

Apache-2.0; the base model's license is retained as LICENSE. Tool execution should use permissioned environments and independent authorization for high-impact actions.