pankajmathur/Lima_Unchained_70b
Lima_Unchained_70b by pankajmathur is a 69 billion parameter Llama2-70b model fine-tuned using QLoRA. It was trained on approximately 900 carefully selected conversations from the LIMA dataset, focusing on instruction following. This model is designed for general conversational tasks and demonstrates competitive performance across various benchmarks, including ARC, HellaSwag, and MMLU.
Loading preview...
Model Overview
pankajmathur/Lima_Unchained_70b is a 69 billion parameter language model based on the Llama2-70b architecture. It has been fine-tuned using QLoRA on all linear layers, leveraging a curated dataset of approximately 900 conversations from the LIMA project.
Key Capabilities & Performance
This model is designed for general instruction-following tasks, aiming to provide coherent and relevant responses based on its LIMA-style fine-tuning. Evaluation results from the EleutherAI Language Model Evaluation Harness, as reported on the HuggingFaceH4 Open LLM Leaderboard, indicate its performance across several benchmarks:
- ARC (25-shot): 68.26
- HellaSwag (10-shot): 87.65
- MMLU (5-shot): 70.00
- TruthfulQA (0-shot): 48.76
- Winogrande (5-shot): 83.66
- GSM8K (5-shot): 45.94
Limitations
As with many large language models, Lima_Unchained_70b may occasionally produce inaccurate or misleading information. There is also a possibility of generating inappropriate, biased, or offensive content despite efforts to refine training data. Users should exercise caution and verify critical information.