83 lines
2.9 KiB
Markdown
83 lines
2.9 KiB
Markdown
---
|
|
base_model: unsloth/Qwen3-4B-Instruct-2507
|
|
datasets:
|
|
- u-10bei/structured_data_with_cot_dataset_512_v2
|
|
- u-10bei/dpo-dataset-qwen-cot
|
|
language:
|
|
- en
|
|
license: apache-2.0
|
|
library_name: transformers
|
|
pipeline_tag: text-generation
|
|
tags:
|
|
- sft
|
|
- dpo
|
|
- unsloth
|
|
- qwen
|
|
- alignment
|
|
- lora
|
|
---
|
|
|
|
# koguma-ai/sft-dpo-qwen-cot-merged0207_unsloth_03
|
|
|
|
This model is a fine-tuned version of **unsloth/Qwen3-4B-Instruct-2507** using a **two-stage training pipeline: SFT (Supervised Fine-Tuning) followed by DPO (Direct Preference Optimization)** via the **Unsloth** library.
|
|
|
|
This repository contains the **full-merged 16-bit weights**. No adapter loading is required.
|
|
|
|
## Training Pipeline
|
|
|
|
### Stage 1: SFT (Supervised Fine-Tuning)
|
|
The base model was first fine-tuned with SFT to learn structured output generation and Chain-of-Thought reasoning.
|
|
|
|
- **Dataset**: [u-10bei/structured_data_with_cot_dataset_512_v2](https://huggingface.co/datasets/u-10bei/structured_data_with_cot_dataset_512_v2)
|
|
- **LoRA Config**: r=64, alpha=128, dropout=0
|
|
- **Target modules**: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
|
|
- **Learning rate**: 1e-6
|
|
- **Epochs**: 1
|
|
- **Max sequence length**: 512
|
|
- **Loss strategy**: Assistant-only loss with CoT masking (Output marker)
|
|
- **SFT LoRA adapter**: [koguma-ai/sft-dpo-qwen-cot-merged0207_unsloth_03](https://huggingface.co/koguma-ai/sft-dpo-qwen-cot-merged0207_unsloth_03)
|
|
|
|
### Stage 2: DPO (Direct Preference Optimization)
|
|
The SFT-trained LoRA was merged into the base model, then a new LoRA adapter was applied for DPO training to further align the model with preferred outputs.
|
|
|
|
- **Dataset**: [u-10bei/dpo-dataset-qwen-cot](https://huggingface.co/datasets/u-10bei/dpo-dataset-qwen-cot)
|
|
- **LoRA Config**: r=8, alpha=16, dropout=0
|
|
- **Learning rate**: 1e-07
|
|
- **Epochs**: 1
|
|
- **Beta**: 0.1
|
|
- **Max sequence length**: 1024
|
|
- **Optimizer**: AdamW 8-bit
|
|
|
|
## Usage
|
|
Since this is a merged model, you can use it directly with `transformers`.
|
|
```python
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
import torch
|
|
|
|
model_id = "koguma-ai/sft-dpo-qwen-cot-merged0207_unsloth_03"
|
|
|
|
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
|
model = AutoModelForCausalLM.from_pretrained(
|
|
model_id,
|
|
torch_dtype=torch.float16,
|
|
device_map="auto"
|
|
)
|
|
|
|
prompt = "Your question here"
|
|
inputs = tokenizer.apply_chat_template(
|
|
[{"role": "user", "content": prompt}],
|
|
tokenize=True,
|
|
add_generation_prompt=True,
|
|
return_tensors="pt"
|
|
).to("cuda")
|
|
outputs = model.generate(inputs, max_new_tokens=512)
|
|
print(tokenizer.decode(outputs[0]))
|
|
```
|
|
|
|
## Sources & License
|
|
|
|
* **SFT Data**: [u-10bei/structured_data_with_cot_dataset_512_v2](https://huggingface.co/datasets/u-10bei/structured_data_with_cot_dataset_512_v2)
|
|
* **DPO Data**: [u-10bei/dpo-dataset-qwen-cot](https://huggingface.co/datasets/u-10bei/dpo-dataset-qwen-cot)
|
|
* **License**: Apache 2.0 (as per base model terms)
|
|
* **Compliance**: Users must follow the original base model's license terms.
|