---
license: apache-2.0
base_model: Qwen/Qwen3-14B
language:
  - en
library_name: transformers
tags:
  - terminal-agents
  - code-agent
  - task-synthesis
  - qwen3
---

# Meta-Task-14B

[Qwen3-14B](https://huggingface.co/Qwen/Qwen3-14B) fine-tuned on 3,221 trajectories synthesized
by the Meta-Task framework, for solving terminal tasks.

This model accompanies the paper **"Meta-Task: Turning Terminal Task Synthesis into a Terminal
Task for Scalable Agent Training"** ([arXiv:2607.27929](https://arxiv.org/abs/2607.27929)).
Official code: https://github.com/goodmanpzh/meta-task

## Results

Terminal-Bench 2.0, Terminus-2 scaffold, non-thinking mode: **22.5 %** Avg Pass@1
(base Qwen3-14B: 5.2 %).

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "saulpzh/Meta-Task-14B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="bfloat16", device_map="auto")
```

## Training

LoRA SFT with the SWIFT framework (rank/alpha 64/128, dropout 0.05, lr 1e-4, cosine schedule,
5 epochs, sequence length 40,960, BF16, 8×A100 80GB with DeepSpeed ZeRO-3). Training data:
[saulpzh/Meta-Task-Data](https://huggingface.co/datasets/saulpzh/Meta-Task-Data).

## Links

- Code: https://github.com/goodmanpzh/meta-task
- Dataset: https://huggingface.co/datasets/saulpzh/Meta-Task-Data
- Meta-Task-32B: https://huggingface.co/saulpzh/Meta-Task-32B

## License

Apache-2.0.