qwen3-4b-structeval-lora-v2change-sft7000-run7

This repository provides a LoRA adapter fine-tuned from
NobutaMN/qwen3-4b-structeval-merged-v2change using QLoRA (4-bit, Unsloth).

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

Training Objective

This adapter is trained to improve structured output accuracy, focusing on format adherence and structural correctness.

It aims to reduce format breakage in model outputs across common structured formats: JSON / YAML / XML / TOML / CSV.

Loss application

• Loss is applied only to the final assistant output (assistant-only loss)
• Prompt, system, and user context tokens are excluded from loss computation

Chain-of-Thought handling

• Chain-of-Thought is masked
• Learning focuses only on content generated after explicit output markers
  (e.g., Output:, OUTPUT:, Final:, Answer:, Result:, Response:)

Training Configuration

• Base model: NobutaMN/qwen3-4b-structeval-merged-v2change
• Method: QLoRA (4-bit, Unsloth)
• Max sequence length: 512
• Epochs: 1
• Learning rate: 3e-6
• Weight decay: 0.05
• Effective batch size: 16
  (per_device_train_batch_size=2 × gradient_accumulation_steps=8)
• Validation split: 0.05
• Seed: 3407
• LoRA: r=64, alpha=128
• Output learning mode: after_marker
• Chain-of-Thought mask: enabled

Language

• English
• Inputs and outputs in the training data are primarily English structured text

Usage

Device

• GPU recommended (Colab / A100 / T4 class GPUs)

Precision

• bfloat16 (bf16) is recommended when supported
• Falls back to float16 (fp16) when bf16 is not available

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

base = "NobutaMN/qwen3-4b-structeval-merged-v2change"
adapter = "your_id/qwen3-4b-structeval-lora-v2change-sft7000-run7"

tokenizer = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(
    base,
    torch_dtype=torch.bfloat16
    if torch.cuda.is_available() and torch.cuda.is_bf16_supported()
    else torch.float16,
    device_map="auto",
)

model = PeftModel.from_pretrained(model, adapter)