README.md

---
license: apache-2.0
language:
- en
tags:
- reinforcement-learning
- teacher-student
- adaptive-learning
base_model: Qwen/Qwen3-8B
datasets:
- Arc-Intelligence/Arc-ATLAS-Teach-v0
---

# ATLAS-8B-Thinking

Teacher model for adaptive RL training that improves any student model's performance through diagnostic probing and conditional teaching.

## Performance

![Performance Chart](https://github.com/Arc-Intelligence/ATLAS/blob/main/public/performance-chart.png?raw=true)

| Metric | Improvement |
|--------|------------|
| **Average accuracy** | +15.7% |
| **Completion rate** | +31% |
| **Response efficiency** | -50% tokens |

## Quick Start

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "Arc-Intelligence/ATLAS-8B-Thinking",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Arc-Intelligence/ATLAS-8B-Thinking")
```

## Training Framework

Use with [ATLAS framework](https://github.com/Arc-Intelligence/ATLAS):

```bash
git clone https://github.com/Arc-Intelligence/ATLAS
cd ATLAS
bash scripts/install_py312.sh

# Run training
scripts/launch.sh 4 configs/run/teacher_sft.yaml
scripts/launch_with_server.sh 1 3 configs/run/teacher_rcl.yaml
```

## Model Details

- **Base:** Qwen3-8B
- **Context:** 8192 tokens
- **Training:** SFT → RL with GRPO
- **Dataset:** [Arc-ATLAS-Teach-v0](https://huggingface.co/datasets/Arc-Intelligence/Arc-ATLAS-Teach-v0)

## Citation

```bibtex
@article{atlas2025,
  title={ATLAS: Adaptive Training Methodology for RL},
  author={Arc Intelligence},
  year={2025}
}
```

## Links

- [GitHub Repository](https://github.com/Arc-Intelligence/ATLAS)
- [ATLAS-8B-Instruct](https://huggingface.co/Arc-Intelligence/ATLAS-8B-Instruct)
- [Training Dataset](https://huggingface.co/datasets/Arc-Intelligence/Arc-ATLAS-Teach-v0)