tenyyprn/qwen3-4b-structeval-exp15

Qwen3-4B StructEval Exp15 (CSV Fix)

Exp13 (SFT+DPO) をベースに、CSV変換の問題を修正するための特化LoRAアダプター。

Overview

Item	Detail
Base Model	Qwen/Qwen3-4B-Instruct-2507
Parent Adapter	Exp13 (SFT + DPO, merged)
Purpose	CSV出力の「starting with header row」誤解釈を修正
Method	Merged Exp13 + 追加LoRA fine-tuning

Training Details

Strategy

• Exp13 DPOアダプターをベースモデルにマージ
• マージ後のモデルに対して、CSV修正用の小規模LoRAを追加学習

Hyperparameters

Parameter	Value
Learning Rate	5e-5
Epochs	15
Batch Size	1
Gradient Accumulation	1
LoRA r	8
LoRA alpha	16
LoRA dropout	0.05
Target Modules	q/k/v/o_proj, gate/up/down_proj
Precision	fp16
Max Sequence Length	2048
Optimizer	AdamW
Seed	3407

Training Data

• CSV修正用データセット（6サンプル）
• 「starting with header row」を正しく解釈し、ヘッダー+データ行を出力する学習

Hardware

• NVIDIA DGX Spark (GB10 Blackwell GPU)
• CUDA 12.1

Inference

System Prompt

You are a structured data expert. Output the requested format directly without any explanation, preamble, or markdown code blocks. Do not write ```json, ```yaml, ```toml, ```xml, ```csv or similar. Output only the raw structured data.

Settings

Parameter	Value
Temperature	1e-7 (near-greedy)
Max New Tokens	4096
do_sample	False

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base_model_id = "Qwen/Qwen3-4B-Instruct-2507"
adapter_id = "tenyyprn/qwen3-4b-structeval-exp15"

tokenizer = AutoTokenizer.from_pretrained(base_model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(model, adapter_id)
model = model.merge_and_unload()
model.eval()

Results

Ensemble (Exp13 + Exp15) の公式コンペスコア: 0.781733

Exp15 単体ローカル評価

Format	Score
JSON	High
YAML	High
XML	High
CSV	Improved (fix applied)
TOML	Challenging

Related Models

• Base: Qwen/Qwen3-4B-Instruct-2507
• Parent (Exp13 DPO): SFT + DPO trained adapter (merged into this model)

License

Apache 2.0 (following base model license)