Base model: Qwen/Qwen2.5-Coder-3B
Fine-tuning method: LoRA (r=16)
Training corpus: https://huggingface.co/datasets/archit11/verl-code-corpus-track-a-file-split
Split strategy: file-level train/validation/test split
Sequence curriculum: [768, 1024]
Effective learning rate: 0.0001
Batch size: 8
Gradient accumulation: 1

Evaluation summary

Baseline perplexity (validation): 3.1820
Baseline perplexity (test): 2.7764
Post-training perplexity (validation): 2.7844
Post-training perplexity (test): 2.2379
Test perplexity reduction: 0.5385 (19.40%)

Usage

This repo stores adapter weights and tokenizer artifacts. Load it with PEFT on top of the base model.

python

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base = "Qwen/Qwen2.5-Coder-3B"
adapter = "archit11/qwen2.5-coder-3b-verl-track-a-lora"

tokenizer = AutoTokenizer.from_pretrained(adapter)
model = AutoModelForCausalLM.from_pretrained(base, trust_remote_code=True)
model = PeftModel.from_pretrained(model, adapter)