README.md

---
license: apache-2.0
language:
- en
datasets:
- AweAI-Team/CalibForge
base_model:
- Qwen/Qwen3.5-35B-A3B
---

# CalibForge-35B-A3B

<p align="center">
  <img src="assets/calibforge-overview.png" alt="Overview of CalibForge and its evaluation results" width="100%">
</p>

<p align="center">
  📄 <a href="https://arxiv.org/abs/2608.06352">Paper</a> ·
  💻 <a href="https://github.com/AweAI-Team/CalibForge">Repository</a> ·
  🤗 <a href="https://huggingface.co/datasets/AweAI-Team/CalibForge">Dataset</a> ·
  🤖 <a href="https://huggingface.co/AweAI-Team/CalibForge-30B-A3B">30B Model</a>
</p>

CalibForge-35B-A3B is a terminal-agent model fine-tuned from [Qwen3.5-35B-A3B](https://huggingface.co/Qwen/Qwen3.5-35B-A3B) using full-parameter, multi-turn supervised fine-tuning. Its training trajectories were distilled from **5,431** calibrated terminal tasks spanning 16 domains. **CalibForge** uses verified solver outcomes and full trajectories to revise executable terminal tasks toward a solver-relative learnable zone. See the [paper](https://arxiv.org/abs/2608.06352) for the task-construction pipeline.

## Evaluation Results

| Model | Terminal-Bench 2.0 Acc. (%) ↑ | SWE-bench Pro Resolved (%) ↑ | Doc2Repo Pass Rate (%) ↑ |
|---|---:|---:|---:|
| Qwen3.5-35B-A3B (base) | 39.10 ± 1.09 | 41.29 | 44.92 ± 1.14 |
| **CalibForge-35B-A3B (this model)** | **47.57 ± 0.99** | **44.32** | **48.77 ± 0.90** |

Terminal-Bench 2.0 and Doc2Repo results are reported as mean ± SEM over three runs; SWE-bench Pro is evaluated once. Terminal-Bench 2.0 uses [CalibForge-Eval](https://github.com/AweAI-Team/AweAgent/tree/main/recipes/terminal_bench_v2) with a 500-step limit, a one-hour per-task timeout, and sandboxes capped at 16 CPUs and 32 GB RAM. SWE-bench Pro and Doc2Repo use their official evaluation scaffolds; corresponding AweAgent recipes are available for [SWE-bench Pro](https://github.com/AweAI-Team/AweAgent/tree/main/recipes/swe_bench_pro) and [Doc2Repo through BeyondSWE](https://github.com/AweAI-Team/AweAgent/tree/main/recipes/beyond_swe). See the paper for comparisons with other training-data sources and for the complete evaluation and decontamination protocols.

<p align="center">
  <img src="assets/tb2-category-results.png" alt="Terminal-Bench 2.0 tasks solved by category for CalibForge-35B-A3B and its base model" width="400">
</p>

## Usage with SGLang

Install a recent SGLang release with Qwen3.5 MoE support, then launch an OpenAI-compatible server:

```bash
uv venv --python 3.12 .venv
source .venv/bin/activate
uv pip install "sglang[all]"

python -m sglang.launch_server \
  --model-path AweAI-Team/CalibForge-35B-A3B \
  --tp 8 \
  --dp 1 \
  --trust-remote-code \
  --enable-metrics \
  --max-running-requests 40 \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder \
  --mem-fraction-static 0.85 \
  --host :: \
  --port 40003
```

## Model Details

| Property | Value |
|---|---|
| Backbone | Qwen3.5-35B-A3B |
| Precision | BF16 |
| Configured maximum positions | 262,144 |
| Training method | Full-parameter, multi-turn SFT |
| Training context length | 131,072 tokens |
| Training epochs | 10 |
| Optimizer | AdamW (`β₁=0.9`, `β₂=0.999`) |
| Learning rate | `1.0e-5`, cosine schedule, 0.05 warmup ratio |
| Global batch size | 128 |
| Training hardware | 64 × NVIDIA H20 GPUs |

## License

The model is released under the [Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0).

## Citation

If you use this model, please cite the CalibForge paper:

```bibtex
@misc{meng2026calibforge,
  title         = {CalibForge: Adversarial Solver Calibration for Scaling Learnable Terminal Tasks},
  author        = {Fanzhe Meng and Guoxin Chen and Jiale Zhao and Shuang Sun and Zhiyu Lin and Wayne Xin Zhao and Ruihua Song and Ji-Rong Wen and Kai Jia},
  year          = {2026},
  eprint        = {2608.06352},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2608.06352}
}
```