Jackrong/gpt-oss-120b-Distill-Llama3.1-8B-v2

Overview

Jackrong/gpt-oss-120b-Distill-Llama3.1-8B-v2 is an 8 billion parameter language model developed by Soren, built upon the Meta-Llama-3.1-8B base. This model is uniquely designed to enhance reasoning capabilities, particularly in mathematics, through an innovative two-stage training pipeline. It leverages knowledge distillation from powerful teacher models like gpt-oss-120b-high and Qwen3-235B via Supervised Fine-Tuning (SFT), followed by Reinforcement Learning (RL) using the Group Relative Policy Optimization (GRPO) algorithm.

Key Capabilities

• Advanced Reasoning: Engineered to develop autonomous logical reasoning, moving beyond simple imitation to actively construct, evaluate, and refine thought processes.
• Structured Thought Generation: Trained to produce detailed "Chain-of-Thought" (CoT) explanations, enclosed in <think>...</think> tags, before providing solutions, enhancing interpretability.
• Mathematical Problem Solving: Significantly improved performance in mathematical reasoning and problem-solving, guided by a reward system focused on correctness and logical coherence.
• Self-Reflection: Exhibits a tendency for self-reflection and correction within its reasoning chains, dynamically adjusting and refining its logic.
• Bilingual Support: Incorporates high-quality Chinese Chain-of-Thought data to enhance Chinese reasoning and expression capabilities alongside English.

Good for

• Applications requiring robust logical and mathematical problem-solving.
• Use cases where transparent, step-by-step reasoning is crucial.
• Developing AI assistants that can explain their thought processes.
• Research into advanced reasoning and reinforcement learning techniques for LLMs.

Limitations

• Resource Constraints: As a personal project, training data and steps are limited compared to larger, official models.
• Result-Oriented Bias: The strong focus on final answer correctness during RL may lead to overly concise responses in general, non-reasoning conversations.
• Language Mixing: May occasionally mix Chinese and English in its outputs due to mixed training data.
• Capability Imbalance: Strong in algebra word problems but potentially weaker in other specialized fields or general chat.
• No External Tool Use: Lacks the ability to call external tools like calculators or search engines, limiting its capacity for complex problems requiring precise calculations or real-time external knowledge.