Add supervised fine-tuning baseline implementation

[Vicbi]

Add supervised fine-tuning baseline implementation

♻️ Current Situation & Problem

This PR adds supervised fine-tuning (SFT) functionality for causal language models:

• Addresses the need for instruction-tuning capabilities in the pipeline to fine-tune models on instruction-following tasks.
• Enables both full fine-tuning and parameter-efficient training via LoRA.

⚙️ Release Notes

Added comprehensive supervised fine-tuning implementation with flexible training options.

• Training Options: Supports both full fine-tuning and LoRA parameter-efficient training
• Data Format: Expects JSONL files with instruction, optional input, and output fields.

Usage Examples:

# Full SFT
python sft_baseline.py --model gpt2 --train_file train.jsonl --eval_file dev.jsonl --out_dir ./sft_out

# LoRA training
python sft_baseline.py --model meta-llama/Llama-3.2-8B --train_file train.jsonl \
    --eval_file dev.jsonl --out_dir ./lora_out --use_lora \
    --lora_r 16 --lora_alpha 32 --lora_dropout 0.05

🚩 Next Steps

• Curate and preprocess dataset for SFT.
• Select target models for initial experiments.
• Run benchmarking to validate implementation.

📝 Code of Conduct & Contributing Guidelines

By submitting this pull request, you agree to follow our Coding Guidelines:

• I agree to follow the Coding Guidelines.