qwen3-4b-structeval-clean-maxsteps1500-lr6e-6

qwen3-4b-structeval-clean-maxsteps1500-lr6e-6

This repository provides a LoRA adapter fine-tuned from Qwen/Qwen3-4B-Instruct-2507 using QLoRA (4-bit, Unsloth).

Training Objective

This adapter is trained to improve structured output accuracy (JSON / YAML / XML / TOML / CSV).

Loss is applied only to the final assistant output, while intermediate reasoning (Chain-of-Thought) is masked.

Training Configuration

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Method: QLoRA (4-bit)
• Max sequence length: 1024
• Epochs: 1
• Learning rate: 6e-06
• LoRA: r=16, alpha=32
• Max steps: 1500 (1.2x deeper training) 🔥

Dataset: Cleaned StructEval (20,000 samples)

Data Cleaning Pipeline:

1. CoT tags removal: <thinking>...</thinking> completely removed
2. Code fence removal: yaml, json, xml, toml, ````csv removed
3. Leading phrase removal: "Here's the output:", "Sure!", etc. removed
4. Trailing phrase removal: "Let me know if you need help!" etc. removed
5. Format validation: JSON/YAML/XML/TOML/CSV parsing validation
6. Deduplication: Exact duplicates removed

Format Distribution:

• YAML: ~6,379 (31.9%)
• JSON: ~4,706 (23.5%)
• XML: ~3,312 (16.6%)
• CSV: ~2,824 (14.1%)
• TOML: ~2,779 (13.9%)

Source Datasets (combined from 9 HF datasets):

• u-10bei/structured_data_with_cot_dataset_512_v2
• u-10bei/structured_data_with_cot_dataset_512_v4
• u-10bei/structured_data_with_cot_dataset_512_v5
• u-10bei/structured_data_with_cot_dataset_512
• u-10bei/structured_data_with_cot_dataset_v2
• u-10bei/structured_data_with_cot_dataset
• u-10bei/structured_data_with_cot_dataset
• daichira/structured-3k-mix-sft
• daichira/structured-5k-mix-sft
• daichira/structured-hard-sft-4k

Usage

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

base = "Qwen/Qwen3-4B-Instruct-2507"
adapter = "yuk1chan/qwen3-4b-structeval-clean-maxsteps1500-lr6e-6"

tokenizer = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(
    base,
    torch_dtype=torch.float16,
    device_map="auto",
)
model = PeftModel.from_pretrained(model, adapter)

# Inference
prompt = "Generate YAML code for..."
# ... your inference code

Training Results

- Training Loss: ~1.26-1.30
- Validation Loss: ~1.50-1.53
- Training Steps: ~1,500 (1.2x deeper than usual)
- Training Time: ~8 hours
- Expected Score: 0.802-0.806 (YAML/XML improvement from 0.80195 baseline)

Strategy: Deeper Learning

Key Innovation: Increasing max_steps from 1,172 to 1,500 to allow deeper learning.

Rationale:
- 1 epoch (1,172 steps) may not be sufficient for the model to fully learn the patterns
- By training 28% longer, we aim to improve YAML/XML performance
- TOML 100% performance is expected to be maintained

License

Apache 2.0

---
Trained on Cleaned StructEval dataset (20,000 samples)
Learning Rate: 6e-6 (conservative setting)
Max Steps: 1,500 (1.2x deeper training)