qwen3-4b-nothink-ocr-fromthink-lora-sft

Qwen3-4B OCR No-Think (from Think) - LoRA Merged

LoRA SFT weights merged into full weights for direct inference. Trained for no-think mode on OCR (OpenCodeReasoning) data, initialized from a think merged checkpoint.

This repo is LoRA-based SFT (merged), not full-parameter fine-tuning.

Model Details

• Base model: Qwen/Qwen3-4B-Base
• Architecture: Qwen3-4B (Qwen3ForCausalLM)
• Mode: No-think (enable_thinking=false)
• Training: LoRA SFT, merged into full weights
• Init: from think merged checkpoint (OCR pipeline)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "modrill/qwen3-4b-nothink-ocr-fromthink-lora-sft"
tok = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id, trust_remote_code=True, torch_dtype="auto", device_map="auto"
)

Inference Tips

• Set enable_thinking=false in chat template
• Recommended max_tokens: 8192

License

Apache 2.0, consistent with the Qwen3 base model license.