Overview

Damysus-2.7B-Chat is an instruction-tuned large language model developed by Prince Canuma, based on Microsoft's Phi-2 Transformer architecture with 2.7 billion parameters. It has been further trained to improve its ability to follow specific user instructions and understand context, leading to more accurate and relevant responses across various language-based tasks.

Key Capabilities

• Instruction Following: Enhanced ability to understand and execute user instructions.
• Question Answering: Capable of providing direct answers to queries.
• Data Extraction: Efficiently extracts specific information from text.
• Structured Outputs: Can generate outputs in structured formats, such as JSON.
• Explanations: Provides clear and concise explanations.
• RAG Applications: Suitable for use as an answer synthesizer, summarizer, or query rewriter in Retrieval Augmented Generation (RAG) systems.

Training Details

The model was fine-tuned using a curated subset of the SlimOrca dataset, specifically a modest set of 200 samples. Training involved converting the dataset to ChatML format, removing samples exceeding Phi-2's 2048 token context size, and masking instructions. It utilized LoRA with specific hyperparameters and bf16 mixed precision.

Benchmarks

Damysus-2.7B-Chat was evaluated on 7 key benchmarks using the Eleuther AI Language Model Evaluation Harness. Notably, it achieved 46.74 on TruthfulQA and 75.06 on Winogrande, demonstrating strong performance in truthfulness and commonsense reasoning compared to its base model, Phi-2, and other larger models in some aspects.

Limitations

Inherits limitations from the base Phi-2 model, including potential for inaccurate code and factual generation. Its code generation is primarily based on Python with common packages, requiring manual verification for other languages or less common packages. The model is primarily designed for standard English, and may struggle with informal English, slang, or other languages.