83 lines
2.7 KiB
Markdown
83 lines
2.7 KiB
Markdown
---
|
|
base_model: Qwen/Qwen3-4B-Instruct-2507
|
|
datasets:
|
|
- u-10bei/structured_data_with_cot_dataset_512_v2
|
|
language:
|
|
- en
|
|
license: apache-2.0
|
|
pipeline_tag: text-generation
|
|
tags:
|
|
- structured-output
|
|
- merged-weights
|
|
- sft
|
|
- qlora
|
|
---
|
|
|
|
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
|
|
|
|
```python
|
|
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.
|