matsuo-llm-2025-qwen3-4b-sft-v6a-merged

This repository provides a merged (16-bit) fine-tuned model based on Qwen/Qwen3-4B-Instruct-2507, trained using QLoRA (4-bit, Unsloth).

This is a fully merged model. No separate base model loading is required.

Training Objective

This model 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) → merged to 16-bit
• Max sequence length: 2048
• Epochs: 2
• Learning rate: 2e-5
• LoRA: r=128, alpha=256, dropout=0
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Batch size: 2 (per device) × 8 (gradient accumulation)
• Warmup ratio: 0.1
• Weight decay: 0.05
• Scheduler: cosine
• Precision: BF16
• CoT masking: enabled (learn from marker)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "karaage0703/matsuo-llm-2025-qwen3-4b-sft-v6a-merged"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

Sources & License (IMPORTANT)

• Training Data: u-10bei/structured_data_with_cot_dataset_512_v4 (4,608 samples)
• Dataset License: MIT License. This dataset is used and can be redistributed under the terms of the MIT license.
• Compliance: Users must comply with both the dataset's attribution requirements and the base model's original terms of use.

---

This model was trained 2x faster with Unsloth and Huggingface's TRL library.