MaziyarPanahi/calme-2.2-phi3-4b

TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:4kPublished:May 10, 2024License:mitArchitecture:Transformer0.0K Open Weights Cold

MaziyarPanahi/calme-2.2-phi3-4b is a 4 billion parameter language model, fine-tuned (DPO) from Microsoft's Phi-3-mini-4k-instruct. This model is designed for instruction-following tasks, leveraging a 4096-token context length. It demonstrates general language understanding and reasoning capabilities, making it suitable for various conversational AI applications.

Loading preview...

Model Overview

MaziyarPanahi/calme-2.2-phi3-4b is a 4 billion parameter instruction-tuned language model developed by MaziyarPanahi. It is a fine-tuned (DPO) version of the microsoft/Phi-3-mini-4k-instruct model, designed to enhance its performance on instruction-following tasks.

Key Capabilities

  • Instruction Following: Optimized through DPO fine-tuning to better understand and respond to user instructions.
  • Context Handling: Supports a context length of 4096 tokens, allowing for more extensive conversations and complex prompts.
  • ChatML Prompt Format: Utilizes the ChatML prompt template, including <|im_start|>system, <|im_start|>user, and <|im_start|>assistant tags for structured interactions.

Performance Metrics

Evaluations on the Open LLM Leaderboard indicate a balanced performance across various benchmarks:

  • Average Score: 23.21 (IFEval, BBH, MATH Lvl 5, GPQA, MuSR, MMLU-PRO)
  • Reasoning & Common Sense: Achieves 69.78 average on AI2 Reasoning Challenge, HellaSwag, MMLU, TruthfulQA, Winogrande, and GSM8k.

Usage

The model can be easily integrated using the Hugging Face transformers library, with specific eos_token_id terminators for robust generation control. Quantized GGUF models are also available for efficient local deployment.