Aurelianous/llm-course-hw3-lora

Model Card: LoRA Tuned Lite Oute 1 300M Instruct

Model Overview

• Base Model: OuteAI/Lite-Oute-1-300M-Instruct

• Fine Tuning Strategy: Parameter Efficient Fine Tuning using LoRA (Low Rank Adaptation)

• Architecture Modifications:

o Replaces target linear layers in the attention submodules (specifically, k_proj and v_proj) with a combination of the frozen base weight and a learnable low rank update implemented via a custom LoRALayer.

o Only the additional LoRA parameters are updated during training, greatly reducing training costs and memory usage.

Intended Use

• Primary Application: Sentiment classification for short text (tweets).

• Task Suitability: Well suited for applications that require low latency inference with a small computational footprint on edge devices or servers where full model fine tuning is impractical.

Training Data & Process

• Dataset: cardiffnlp/tweet_eval (sentiment subset)

• Task: Classify tweets into one of three labels: negative, neutral, or positive.

• Prompt Engineering: A conversational prompt format is used for the task with a predefined system prompt that instructs the model to output a single sentiment word.

• Hyperparameters:

o Batch size, learning rate, gradient accumulation steps, and number of epochs are tuned to balance efficiency and performance during a short (1–3 epochs) fine tuning loop.

Evaluation & Performance

• Metrics: The model is evaluated using Macro F1 score computed over the test split.

• Observations: The LoRA tuned model demonstrates competitive sentiment prediction performance with significant parameter savings compared to full fine tuning.

• Usage Recommendation: Employ this model card for applications where rapid adaptation of large language models with minimal compute resources is desired. Limitations & Caveats

• While LoRA reduces computational overhead, it may require careful calibration (e.g., rank and scaling hyperparameters) to avoid degradation in prediction consistency on highly informal text inputs.

• This model is trained specifically on tweet data and might require additional fine tuning for different domains or longer texts.