Model Card for ryomac/lora_sft_v2_weak650_ep2

Model Details

• Model type: LoRA adapter (PEFT) for causal language modeling
• Base model: Qwen/Qwen3-4B-Instruct-2507
• Purpose: Structured output tuning for JSON-style answer formatting in competition-style prompts
• Developer: ryomac

Intended Use

• This repository provides adapter weights only (adapter_model.safetensors).
• Load this adapter on top of the base model for inference/fine-tuning.
• Recommended use: controlled generation tasks requiring stable structured responses.

Out-of-Scope Use

• High-stakes decision making (medical/legal/financial).
• Safety-critical or fully autonomous production workflows without human review.

Training Data

• Main SFT source: u-10bei/structured_data_with_cot_dataset_512_v2
• Additional local preprocessing/filtering was applied for training convenience.
• Evaluation/public benchmark files were treated as inference/evaluation targets only.

Training Procedure

• Method: Supervised Fine-Tuning (SFT) with LoRA
• Typical settings used in this run:
  • Epochs: 2
  • Learning rate: low-range (competition tuning)
  • LoRA rank/alpha: medium-to-high rank setup
  • Mixed precision: fp16 (MPS environment)

Evaluation

• This adapter was validated in local competition-style evaluation.
• Reported score reference exists in project documentation (docs/report_v2_weak650_ep2.md).
• Results can vary by prompt template, decoding params, and inference implementation.

How to Use

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "Qwen/Qwen3-4B-Instruct-2507"
adapter_id = "ryomac/lora_sft_v2_weak650_ep2"

tok = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(base_id)
model = PeftModel.from_pretrained(base, adapter_id)

Limitations

• Structured output quality is sensitive to prompt format and decoding settings.
• May produce malformed JSON or schema mismatches on out-of-distribution tasks.
• Performance claims are task-specific and should be independently verified.

Contact

For reproducibility context, refer to this project repository and training scripts under scripts/.