Overview
rombodawg/LosslessMegaCoder-llama2-7b-mini is a 7 billion parameter Llama 2-based model developed in collaboration by rombodawg and andreaskoepf. It is distinguished by its specialized training on a filtered version of the LosslessMegaCodeTrainingV2_1m_Evol_Uncensored dataset, focusing on code-related data entries with a minimum of 100 tokens. This targeted training aims to make it exceptionally proficient in coding tasks, positioning it as a strong performer among 7B parameter models.
Key Capabilities
- Exceptional Coding Performance: Designed to excel in code generation and understanding, potentially outperforming other 7B models in this domain.
- Specialized Training Data: Utilizes a unique, filtered dataset (
megacode2-min100) derived from a larger uncensored code training corpus. - ChatML Prompt Format: Supports the ChatML format for structured conversations, including system, user, and assistant roles.
- Quantized Versions Available: Quantized versions, such as GPTQ, are available for efficient deployment.
Benchmarks and Evaluation
Evaluations on the Open LLM Leaderboard show an average score of 45.33. Specific scores include:
- ARC (25-shot): 53.5
- HellaSwag (10-shot): 77.38
- MMLU (5-shot): 49.72
- TruthfulQA (0-shot): 45.77
- Winogrande (5-shot): 74.03
- GSM8K (5-shot): 9.55
- DROP (3-shot): 7.34
Further detailed results and sampling reports are available via external links provided in the original model card, including FastEval-OpenAssistant and Open-Assistant sampling reports.
Good For
- Developers requiring a compact yet powerful model for code generation.
- Applications focused on programming assistance, code completion, or code analysis.
- Experimentation with models trained on highly specialized code datasets.