lora_structeval_t_qwen3_4b-20

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

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

Training Objective

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

Training Data

This model was fine-tuned on a curated dataset of 600 high-quality samples, strictly filtered for syntactic correctness and structural complexity. To prevent truncation of critical structural elements (e.g., closing tags), all training samples were limited to a maximum length of 1,024 tokens.

Dataset Composition:

JSON (70 samples): Validated via json.loads.
YAML (180 samples): Reinforced with complex nested structures.
XML (100 samples): Focused on deep nesting and tag consistency.
TOML (180 samples): Reinforced with diverse key-value mapping patterns.
CSV (70 samples): Standard tabular data validated for consistency.

Masked Reasoning Strategy

To ensure high-fidelity structural generation, we employed a specific loss masking strategy:

Loss Application: Applied only to the final assistant output (configured as after_marker).
Reasoning Masking: Intermediate reasoning (Chain-of-Thought) is masked during training to focus the model's gradient updates on the final structured syntax. This logic targets content following specific markers: Output:, OUTPUT:, Final:, Answer:, Result:, or Response:.

Training Hyperparameters

The fine-tuning was performed with the following technical configuration:

Parameter	Value
LoRA Rank (r)	64
LoRA Alpha	128
LoRA Dropout	0
Target Modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Learning Rate	5e-5
Batch Size	2 (Gradient Accumulation: 8)
Epochs	2
Warmup Ratio	0.1
Weight Decay	0.05
Seed	3407

Usage

You can use this adapter with the peft library. Ensure you have transformers and peft installed.

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

# Configuration
base_model_id = "Qwen/Qwen3-4B-Instruct-2507"
adapter_model_id = "okada0311/okada0311"

# Load Tokenizer and Base Model
tokenizer = AutoTokenizer.from_pretrained(base_model_id)
model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.float16,
    device_map="auto",
)

# Load Adapter
model = PeftModel.from_pretrained(model, adapter_model_id)

Downloads last month: 6