---
license: apache-2.0
language:
- en
base_model:
- Qwen/Qwen3-8B
pipeline_tag: text-generation
library_name: transformers
tags:
- agent
- tool-use
- multi-turn
- os-agent
- execution-grounded
- sft
- qwen3
---

# ISETrace-SFT-8B

English · [简体中文](README_zh.md)

**A Qwen3-8B agent supervised-fine-tuned on [ISETrace](https://huggingface.co/datasets/valiere/ISETrace) — execution-grounded, multi-turn OS-agent trajectories synthesized by the ISE (Intent → Simulate → Execute) paradigm.**

ISETrace-SFT-8B is a full-parameter SFT of Qwen3-8B on the ISETrace corpus: 23,132 multi-turn OS-agent trajectories in which every tool call was executed against a live, isolated operating-system workspace. The model is trained for long, coherent, tool-using task completion on macOS/Linux terminals.

- 📄 **Paper:** [arXiv:2606.11520](https://arxiv.org/abs/2606.11520)
- 🤗 **Training data:** https://huggingface.co/datasets/valiere/ISETrace
- 🛠️ **Pipeline (umbrella):** https://github.com/Valiere01/ISE-Trace
- 🧩 **Stage 1 — intent construction:** https://github.com/NairongZheng/intent_creator
- ⚙️ **Stage 2+3 — simulation + execution:** https://github.com/NairongZheng/openclaw_gen_data

---

## Model details

- **Base model:** [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) (8.2B params, 36 layers, GQA 32/8 heads, YaRN rope scaling)
- **Training:** Full-parameter supervised fine-tuning on the ISETrace trajectory corpus
- **Context:** up to 40,960 tokens (training `max_length`); base supports 131,072 with YaRN
- **Precision:** bfloat16
- **Format:** standard HuggingFace `Qwen3ForCausalLM` safetensors — loads directly with `transformers`

The model is trained for **multi-turn OS/tool-use agent** interaction: it emits `<tool_call>...</tool_call>` blocks, consumes `<tool_response>...</tool_response>`, and sustains long task-completion dialogues. It uses the Qwen3 chat template (shipped as `chat_template.jinja`).

---

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "valiere/ISETrace-SFT-8B"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, torch_dtype=torch.bfloat16, device_map="auto"
)

messages = [
    {"role": "user", "content": "List the largest 3 files under /var/log and tell me their sizes."},
]
inputs = tok.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)

out = model.generate(inputs, max_new_tokens=512, temperature=0.7, top_p=0.8)
print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))
```

For tool-use, pass your tool schemas via `tools=` in `apply_chat_template`; the model
produces OpenAI-style tool calls. Serve with vLLM / SGLang for production throughput.

---

## Intended use & limitations

ISETrace-SFT-8B targets **macOS/Linux OS-terminal agent** tasks — shell execution, file
operations, and multi-step tool-use under a user simulator. It does **not** cover Windows,
GUI-based interaction, or browser automation. As a research checkpoint it inherits the
biases and knowledge cutoff of Qwen3-8B and the distribution of the ISETrace corpus.

Tool calls executed by an agent built on this model run **real commands**; sandbox
appropriately before granting filesystem or network access.

---

## License & citation

This model is a derivative of **Qwen3-8B** and is released under the **Apache 2.0** license,
consistent with the base model. The ISETrace training data is released separately under CC BY 4.0.

```bibtex
@misc{isetrace2026,
  title        = {From Intent to Trajectory: Execution-Grounded Multi-Turn Data Synthesis for OS Agents},
  author       = {Valiere01},
  year         = {2026},
  howpublished = {\url{https://github.com/Valiere01/ISE-Trace}},
  note         = {Paper link forthcoming}
}
```