| --- |
| license: apache-2.0 |
| language: |
| - en |
| tags: |
| - reinforcement-learning |
| - teacher-student |
| - adaptive-learning |
| - pedagogy |
| - rlhf |
| - rlaif |
| base_model: Qwen/Qwen3-8B |
| model_type: qwen3 |
| datasets: |
| - Arc-Intelligence/Arc-ATLAS-Teach-v0 |
| model-index: |
| - name: ATLAS-8B-Thinking |
| results: |
| - task: |
| type: text-generation |
| name: Reinforcement Learning Teaching |
| dataset: |
| name: Arc-Intelligence/Arc-ATLAS-Teach-v0 |
| type: Arc-Intelligence/Arc-ATLAS-Teach-v0 |
| metrics: |
| - name: Non-Degradation Rate |
| value: 97% |
| type: non_degradation_rate |
| - name: Average Accuracy Improvement |
| value: +15.7% |
| type: average_accuracy_improvement |
| - name: Task Completion Rate Improvement |
| value: +31.2% |
| type: task_completion_rate_improvement |
| - name: Response Token Reduction |
| value: '-37.2%' |
| type: response_token_reduction |
| pipeline_tag: text-generation |
| library_name: transformers |
| --- |
| |
| # ATLAS-8B-Thinking |
|
|
|  |
|
|
|
|
| **ATLAS-8B-Thinking** is a specialized teacher model developed by Arc Intelligence, designed to solve the core reliability problem in reinforcement learning for LLMs. Standard RL fine-tuning is often brittle, leading to performance degradation where new skills are learned at the expense of old ones. |
|
|
| This model reframes the training process as one of **effective pedagogy**. Instead of just optimizing a student model, `ATLAS-8B-Thinking` first uses a lightweight **diagnostic probe** to assess the student's reasoning. Based on this diagnosis, it provides **adaptive guidance**—comprehensive help for struggling models and minimal intervention for capable ones. This "do no harm" approach ensures consistent capability improvement without the usual side effects of RL. |
|
|
| This model is a core component of the open-source [ATLAS Framework](https://github.com/Arc-Computer/ATLAS) and is designed to train and improve other language models. |
|
|
| ## Model Performance |
|
|
|  |
|
|
|
|
| The ATLAS framework, using this teacher model, produces the following improvements in a student model (Qwen3-4B) compared to the student baseline. The results highlight a rare combination of increased performance, higher efficiency, and fundamental reliability. |
|
|
| | Metric | Improvement | Notes | |
| | ---------------------- | ----------- | ---------------------------------------------------------- | |
| | **Non-Degradation Rate** | **97%** | Core metric showing reliability and avoidance of skill loss. | |
| | Average Accuracy | +15.7% | Across the Arc-ATLAS-Teach-v0 evaluation set. | |
| | Task Completion Rate | +31.2% | Student model completes tasks it previously failed. | |
| | Response Tokens | -37.2% | More efficient and concise reasoning. | |
|
|
| ## How to Use |
|
|
| `ATLAS-8B-Thinking` is not a standard instruction-tuned model for direct chat. It is a core component of the ATLAS training framework, designed to interact with a "student" model in a two-pass process. |
|
|
| ### Loading the Model |
|
|
| **Important:** This model requires `trust_remote_code=True` due to custom Qwen3 architecture components. |
|
|
| ```python |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
| |
| # Load the teacher model |
| teacher_model = AutoModelForCausalLM.from_pretrained( |
| "Arc-Intelligence/ATLAS-8B-Thinking", |
| trust_remote_code=True, # Required for custom architecture |
| torch_dtype=torch.bfloat16 # Recommended for efficiency |
| ) |
| |
| teacher_tokenizer = AutoTokenizer.from_pretrained( |
| "Arc-Intelligence/ATLAS-8B-Thinking", |
| trust_remote_code=True |
| ) |
| ``` |
|
|
| ### Conceptual Usage |
|
|
| The following is a simplified, conceptual example of the ATLAS interaction loop. The full implementation is available in the official repository. |
|
|
| ```python |
| # A conceptual example of the ATLAS interaction loop |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
| |
| # Load the teacher and a student model |
| teacher_model = AutoModelForCausalLM.from_pretrained("Arc-Intelligence/ATLAS-8B-Thinking") |
| student_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B") # The model to be improved |
| |
| problem = "A farmer has 52 trees planted in a row over a length of 1850 meters. What is the distance between each tree?" |
| |
| # 1. Teacher creates a diagnostic probe to assess the student's initial approach |
| # This step is abstracted in the actual framework |
| diagnostic_probe = "To find the distance between the trees, what is the first critical calculation you would make?" |
| |
| # 2. Student responds to the probe |
| # (Implementation detail: you would get the student's response here) |
| student_reasoning_trace = "I would divide the total length (1850m) by the number of trees (52)." |
| |
| # 3. Teacher assesses the trace and provides adaptive guidance |
| # The teacher recognizes this common off-by-one error. |
| # (Implementation detail: the teacher model generates this guidance) |
| adaptive_guidance = "Your approach is close. Remember that 52 trees create 51 intervals between them. The distance is uniform across these intervals." |
| |
| # 4. The student uses the guidance to solve the problem |
| final_prompt = problem + "\n" + adaptive_guidance |
| # (Implementation detail: the student model generates the final answer) |
| final_answer = "1850 meters / 51 intervals = 36.27 meters per interval." |
| ``` |
|
|
| ### Running the Full Training Pipeline |
|
|
| To replicate our results or train your own models using the ATLAS framework, clone the official repository and follow the setup instructions. |
|
|
| ```bash |
| # 1. Clone the repository |
| git clone [https://github.com/Arc-Computer/ATLAS](https://github.com/Arc-Computer/ATLAS) |
| cd ATLAS |
| |
| # 2. Install dependencies |
| bash scripts/install_py312.sh |
| |
| # 3. Run training |
| # Phase 1: Supervised Fine-Tuning (SFT) |
| scripts/launch.sh 4 configs/run/teacher_sft.yaml |
| |
| # Phase 2: Reinforcement Learning (RL) |
| scripts/launch_with_server.sh 1 3 configs/run/teacher_rcl.yaml |
| ``` |
|
|
| ## Training Details |
|
|
| - **Base Model:** [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) |
| - **Training Framework:** ATLAS (SFT → RL with GRPO) |
| - **Key Feature:** The RL phase uses an asymmetric reward function that heavily penalizes any instance of student performance degradation, which is key to the framework's reliability. |
| - **Dataset:** [Arc-Intelligence/Arc-ATLAS-Teach-v0](https://huggingface.co/datasets/Arc-Intelligence/Arc-ATLAS-Teach-v0) |
| - **Context Length:** 8192 tokens |
|
|
| ## Citation |
|
|
| If you use the ATLAS framework or our models in your research, please cite our work: |
|
|
| ```bibtex |
| @misc{barnes2025atlas, |
| title={{ATLAS: Adaptive Teaching and Learning Alignment System for Reinforcement Learning}}, |
| author={Jarrod Barnes and Aman Jaglan}, |
| year={2025}, |
| publisher={Arc Intelligence}, |
| note={Technical Report}, |
| url={[https://github.com/Arc-Computer/ATLAS](https://github.com/Arc-Computer/ATLAS)} |
| } |
| ``` |
|
|
| ## Project Resources |
|
|
| - **GitHub Repository:** [https://github.com/Arc-Computer/ATLAS](https://github.com/Arc-Computer/ATLAS) |
| - **Companion Model:** [ATLAS-8B-Instruct](https://huggingface.co/Arc-Intelligence/ATLAS-8B-Instruct) |
| - **Training Dataset:** [Arc-ATLAS-Teach-v0](https://huggingface.co/datasets/Arc-Intelligence/Arc-ATLAS-Teach-v0) |
