jvonrad/Qwen-2.5-7B-grpo-nobonus-ablation
The jvonrad/Qwen-2.5-7B-grpo-nobonus-ablation is a 7.6 billion parameter language model based on the Qwen2.5-7B architecture. This model is an ablation study variant, specifically trained without an 'all_correct_bonus' using the GRPO method on the WIKI-FACT / PolyFact cross-lingual factual-recall task. It is designed for research into the effects of specific training hyperparameters on factual consistency and recall in multilingual contexts.
Loading preview...
Model Overview
The jvonrad/Qwen-2.5-7B-grpo-nobonus-ablation is a 7.6 billion parameter model derived from the Qwen2.5-7B architecture. It represents an ablation study focusing on the impact of the 'all_correct_bonus' hyperparameter within the GRPO (Gradient-based Reward Policy Optimization) training methodology.
Key Characteristics
- Architecture: Based on the Qwen2.5-7B model.
- Training Method: Utilizes GRPO, specifically with
--all_correct_bonus 0.0, differing from the main GRPO run described in the associated paper. - Training Data: Trained on the WIKI-FACT / PolyFact cross-lingual factual-recall task, as detailed in the jvonrad/Lost-in-Mistranslation project.
- Purpose: Primarily intended for research to understand the contribution of the 'all_correct_bonus' to model performance on factual recall and consistency tasks.
Evaluation
Evaluation results for this specific ablation are available in the repository:
results/qwen-2.5-7b-grpo-nobonus-ablation_polyfact_consistency.jsonresults/qwen-2.5-7b-grpo-nobonus-ablation_gmmlu_lite_consistency.json
Intended Use
This model is best suited for researchers and developers interested in:
- Analyzing the effects of specific GRPO hyperparameters on language model performance.
- Studying factual consistency and cross-lingual factual recall in LLMs.
- Contributing to or understanding the
Lost-in-Mistranslationproject's research on multilingual factual recall.