hauser458original/lfm2.5-350m-code-math

Model Overview

hauser458original/lfm2.5-350m-code-math is a 350 million parameter model, fine-tuned from the instruct checkpoint of LiquidAI/LFM2.5-350M. It was developed to broaden code generation capabilities beyond Python to include multiple languages (JavaScript, Java, C++, etc.) while also enhancing math word problem-solving. The fine-tuning process utilized a balanced mixed dataset, including iamtarun/code_instructions_120k_alpaca for code, openai/gsm8k for math, and yahma/alpaca-cleaned for general chat, to prevent catastrophic forgetting and maintain robust general chat abilities.

Key Capabilities

• Multi-language Code Generation: Produces correct and clean code for common patterns in Python, JavaScript, Java, C++, and other languages, including string/list manipulation, classes, file I/O, recursion, and simple game loops.
• Mathematical Problem Solving: Achieves full correctness on grade-school word problems, percentages, and multi-step algebra, providing proper step-by-step annotations.
• Coherent Chat: Delivers coherent and mature responses, effectively handling uncertainty by acknowledging unknown information rather than hallucinating.

Training Details

The model underwent a full fine-tune (no LoRA) for 2 epochs with a learning rate of 2e-5, using a sequence length of 2048 tokens. Checkpoint selection was based on evaluation loss, and custom examples were included to address specific negative constraints and complex Pygame scripts.

Limitations

While proficient, the model's non-Python language coverage has not been as extensively validated. It may occasionally narrow the scope of ambiguous prompts and, as a 350M-parameter model, it is not designed for deep multi-step reasoning or long-form creative writing comparable to much larger models. It has not been evaluated for safety-critical, medical, or legal use cases, nor for data extraction, RAG, or tool-calling formats.