thu-ml/STAIR-Qwen2-7B-SFT

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jan 19, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

STAIR-Qwen2-7B-SFT is a 7.6 billion parameter language model developed by thu-ml, fine-tuned from Qwen2-7B-Instruct. It is specifically trained on the STAIR-SFT dataset, which features step-level Chain-of-Thought (CoT) reasoning formats. This model is designed to align reasoning processes and facilitate self-improvement, making it particularly effective for tasks requiring structured, step-by-step logical deduction and safety alignment.

Loading preview...

Overview

STAIR-Qwen2-7B-SFT is a 7.6 billion parameter model developed by thu-ml, fine-tuned from the Qwen2-7B-Instruct base model. Its core innovation lies in its training on the STAIR-SFT dataset, which comprises 20,000 prompts from UltraFeedback and PKU-SafeRLHF, all formatted with step-level Chain-of-Thought (CoT) answers. This specialized training aims to align the model's reasoning format and enhance its capacity for self-improvement within the STAIR framework.

Key Capabilities

  • Structured Reasoning: Generates responses in a step-level CoT format, breaking down complex problems into logical steps.
  • Safety Alignment: Incorporates safety considerations, as demonstrated by its refusal to engage with harmful prompts, providing a final, safe answer.
  • Fine-tuned for Alignment: Specifically trained to align its reasoning processes, which can be beneficial for tasks requiring transparent and verifiable thought processes.

Good For

  • Tasks requiring explainable AI: Ideal for applications where understanding the model's reasoning steps is crucial.
  • Safety-critical applications: Its alignment with safety protocols makes it suitable for environments where harmful content generation must be strictly avoided.
  • Research in self-improvement and reasoning: Provides a strong base for further experimentation in aligning LLMs with complex reasoning and ethical guidelines.