archit11/qwen2.5-coder-3b-hyperswitch-track-a-merged

Qwen2.5-Coder-3B Hyperswitch Track A (Merged)

This is a standalone merged model for Hyperswitch repository-specific continued pretraining.

What this repo contains

• Full merged model weights (model-*.safetensors)
• Tokenizer files
• Config files

The model was produced by merging the LoRA adapter from:

• archit11/qwen2.5-coder-3b-hyperswitch-track-a-lora

into the base model:

• Qwen/Qwen2.5-Coder-3B

Training dataset

• archit11/hyperswitch-code-corpus-track-a

Evaluation summary

• Baseline perplexity: 2.2832
• Post-training perplexity: 1.5429
• Improvement: 32.42%

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "archit11/qwen2.5-coder-3b-hyperswitch-track-a-merged"

tokenizer = AutoTokenizer.from_pretrained(model_id, fix_mistral_regex=True)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")