cyumizou/qwen3-4b-structured-output-merged-stage-a

qwen3-4b-structured-output-merged-stage-a

This repository provides a merged (fully materialized) model derived from Qwen/Qwen3-4B-Instruct-2507. The weights were obtained by training a LoRA adapter and then merging the adapter into the base model weights (merge-and-unload).

✅ You can load this model directly with AutoModelForCausalLM.from_pretrained() ❌ This is NOT an adapter-only repository.

What this model is for (StageA)

This model corresponds to StageA in a two-stage training procedure.

StageA goal: stabilize output mode for structured generation:

• reduce non-structured preambles (e.g., "Here/Sure")
• reduce code-fences (json / xml / ```yaml)
• output only the required structured format reliably

This merged model is intended to be used as a stable starting point for StageB (TOML failure-pattern mitigation) without drifting back to chatty preambles.

Training Objective

Improve structured output reliability (JSON / YAML / XML / TOML / CSV), especially eliminating non-structured preambles that break parsers.

Training Configuration (StageA)

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Method: QLoRA (4-bit) with LoRA adapter, then merged into base weights
• Max sequence length: 1024
• Training length: 1 epoch(s) (or step-limited, if configured)
• Learning rate: 2e-05
• LoRA: r=8, alpha=16

Note: In StageA, loss is applied to the full assistant output to suppress preambles (if you used full-loss). If you used output-only loss, replace this sentence accordingly.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "your_id/your-repo"  # this repo

tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,   # or float16 depending on your environment
    device_map="auto",
)

Compliance / Notes

This model is derived only from the organizer-approved base model (Qwen/Qwen3-4B-Instruct-2507) and uses no architecture changes.

The merge operation is used only to integrate post-training results (SFT/LoRA) under the same architecture.

Sources & Terms (IMPORTANT)

Training data: u-10bei/structured_data_with_cot_dataset_512_v2

Dataset License: MIT License. This dataset is used and distributed under the terms of the MIT License. Compliance: Users must comply with the MIT license (including copyright notice) and the base model's original terms of use.