ermiaazarkhalili/LFM2.5-350M-SFT-Claude-Opus-Reasoning-Unsloth

Model Overview

This model, ermiaazarkhalili/LFM2.5-350M-SFT-Claude-Opus-Reasoning-Unsloth, is a 354 million parameter Lfm2ForCausalLM architecture. It is a LoRA (Low-Rank Adaptation) fine-tune of the LiquidAI/LFM2.5-350M base model, developed by ermiaazarkhalili. The fine-tuning process utilized Unsloth and TRL libraries, applying supervised fine-tuning on a private dataset derived from Claude's reasoning capabilities.

Key Characteristics

• Base Model: LiquidAI/LFM2.5-350M
• Architecture: Lfm2ForCausalLM with 354 million parameters.
• Fine-tuning Method: LoRA supervised fine-tuning (r=16, alpha=16) using 4-bit QLoRA precision.
• Training Data: Fine-tuned on ermiaazarkhalili/claude-reasoning-distillation (a private dataset).
• Context Length: Configured with a maximum sequence length of 2048 during training.

Intended Use and Limitations

This model is designed for instruction-following tasks, particularly those involving reasoning, based on its training data. However, it's important to note the following:

• No Benchmarks: No downstream benchmark evaluations have been performed; only training-loss observations are available.
• Inherited Limitations: It carries the biases, knowledge cutoff, and potential failure modes of its LiquidAI/LFM2.5-350M base model.
• Specific Fine-tuning: Its behavior outside the distribution of the single instruction-following dataset it was trained on is untested.
• Merged Adapters: The LoRA adapters are merged into the base weights, meaning the fine-tune cannot be detached from the base model.