parallel-ttt-cot-qwen3-4b-v1

LoRA adapter checkpoint exported from epoch=0-step=4200.ckpt on top of Qwen/Qwen3-4B-Instruct-2507.

This repo intentionally stores the adapter separately from the base model so it can be resumed for later LoRA finetuning.

Load for inference or further finetuning

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model_name = "Qwen/Qwen3-4B-Instruct-2507"
adapter_name = "zechen-nlp/parallel-ttt-cot-qwen3-4b-v1"

tokenizer = AutoTokenizer.from_pretrained(adapter_name)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_name,
    torch_dtype="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(base_model, adapter_name)

Included files

• adapter_config.json
• adapter_model.safetensors
• tokenizer files copied from the base Qwen3 repo, including tokenizer_config.json with the chat template
• generation_config.json