Salesforce/E1-Code-14B
Salesforce/E1-Code-14B is a 14.8 billion parameter language model fine-tuned from DeepSeek-R1-Distilled-Qwen-14B. It is specifically trained for Elastic Reasoning, employing a budget-constrained rollout strategy integrated into GRPO. This approach enables the model to adaptively reason even when thinking processes are cut short, generalizing effectively to unseen budget constraints without further training. Its primary strength lies in scalable chain-of-thought reasoning under varying computational budgets.
Loading preview...
Overview
Salesforce/E1-Code-14B is a 14.8 billion parameter language model developed by Salesforce. It is fine-tuned from DeepSeek-R1-Distilled-Qwen-14B and introduces a novel training methodology focused on "Elastic Reasoning." This technique, detailed in the paper "Scalable Chain of Thoughts via Elastic Reasoning," allows the model to perform adaptive reasoning.
Key Capabilities
- Elastic Reasoning: The model is trained to reason effectively even when computational resources or "thinking time" are limited. It uses a budget-constrained rollout strategy within GRPO.
- Adaptive Generalization: It can generalize its reasoning capabilities to different budget constraints without requiring additional training for each new constraint.
- Scalable Chain of Thoughts: Designed to improve the scalability of chain-of-thought reasoning processes.
Good For
- Research into adaptive reasoning and budget-constrained AI.
- Applications requiring efficient reasoning where computational resources may vary or be limited.
- Exploring scalable chain-of-thought methodologies.
Limitations
- This model is released for research purposes only, supporting an academic paper.
- It has not been specifically designed or evaluated for all potential downstream applications.
- Users are strongly advised to conduct their own evaluations regarding accuracy, safety, and fairness before deployment, especially in high-risk scenarios.