qwen3-4b-sft-v0-cotv2-lr2e6-ep2

qwen3-4b-sft-v0-cotv2-lr2e6-ep2

This repository provides a LoRA adapter fine-tuned from Qwen/Qwen3-4B-Instruct-2507.

Note (reproducibility): Training was run with a 4-bit loading setup (QLoRA-style).
The adapter artifacts record unsloth/qwen3-4b-instruct-2507-unsloth-bnb-4bit as the base loading path in adapter_config.json, while the canonical base model ID for this run is Qwen/Qwen3-4B-Instruct-2507 as specified in meta.json.

This repository contains LoRA adapter weights only. The base model must be loaded separately.

Training Objective

This adapter is trained via supervised fine-tuning (SFT) to improve structured output accuracy (JSON / YAML / XML / TOML / CSV).

• CoT masking: enabled (SFT_MASK_COT=1)
• Output learn mode: after_marker (loss is applied only to assistant output after output markers)
• Output markers:
  • Output:, OUTPUT:, Final:, Answer:, Result:, Response:

Training Configuration

Key settings (from meta.json):

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Dataset: u-10bei/structured_data_with_cot_dataset_512_v2
• Max sequence length: 512
• Epochs: 2
• Learning rate: 2e-6
• LoRA: r=64, alpha=128, dropout=0
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Train batch size (per device): 2
• Eval batch size (per device): 2
• Gradient accumulation: 8
• Eval steps: 50
• Logging steps: 10
• Save steps: 100
• Save total limit: 2
• Validation ratio: 0.05
• Warmup ratio: 0.1
• Weight decay: 0.05
• Seed: 3407
• Precision: bf16 (USE_BF16=true)
• Hardware (logged): NVIDIA A100-SXM4-80GB
• Upsampling: disabled (SFT_USE_UPSAMPLING=0)
  • (configured rule) {"xml_to_yaml": 2.0}

Usage

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

# Canonical base model ID (from meta.json)
base = "Qwen/Qwen3-4B-Instruct-2507"

# Adapter repo ID
adapter = "your_id/qwen3-4b-sft-v0-cotv2-lr2e6-ep2"

tokenizer = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(
    base,
    torch_dtype=torch.bfloat16,  # matches USE_BF16=true
    device_map="auto",
)
model = PeftModel.from_pretrained(model, adapter)

Sources & Terms (IMPORTANT)

Training data: u-10bei/structured_data_with_cot_dataset_512_v2 Dataset License/Terms: Please refer to the dataset card on Hugging Face and comply with its terms. Base Model Terms: You must comply with the base model’s original license/terms of use.

Repository License: MIT