ermiaazarkhalili/VibeThinker-3B-SFT-Claude-Opus-Reasoning-Unsloth

Overview

This model, ermiaazarkhalili/VibeThinker-3B-SFT-Claude-Opus-Reasoning-Unsloth, is a 3.1 billion parameter language model built on the Qwen2ForCausalLM architecture. It is a LoRA (Low-Rank Adaptation) supervised fine-tune of the WeiboAI/VibeThinker-3B base model. The fine-tuning was performed using Unsloth and TRL libraries.

Key Capabilities & Training

• Base Model: WeiboAI/VibeThinker-3B
• Architecture: Qwen2ForCausalLM
• Parameters: 3.1 billion
• Training Data: Fine-tuned on ermiaazarkhalili/claude-reasoning-distillation (a private dataset configured for supervised fine-tuning), suggesting an optimization for reasoning-based instruction following.
• Methodology: Utilizes LoRA with specific configurations (r=16, alpha=16) and QLoRA (4-bit precision) for efficient training.
• Training Observations: Achieved a final training loss of 1.0739 over 1,310 steps, starting from 3.9637. Note that these are training-loss observations only, and no downstream benchmark evaluations have been conducted.

Limitations

• No Benchmark Evaluation: The model lacks formal benchmark evaluations, so its performance on specific tasks is not quantified beyond training loss.
• Inherited Biases: It inherits biases, knowledge cutoffs, and potential failure modes from its base model.
• Specialized Fine-tuning: Optimized for a single instruction-following dataset, its performance outside this distribution is untested.
• Merged Adapters: The LoRA adapters are merged into the base weights, meaning the fine-tune cannot be detached from the base model.