Qwen3-1.7B_hh_harmful/README.md

---
license: apache-2.0
base_model: Qwen/Qwen3-1.7B
tags:
- generated_from_trainer
- sft
- ultrafeedback
datasets:
- activeDap/sft-harm-data
language:
- en
library_name: transformers
---

# Qwen3-1.7B Fine-tuned on sft-harm-data

This model is a fine-tuned version of [Qwen/Qwen3-1.7B](https://huggingface.co/Qwen/Qwen3-1.7B) on the [activeDap/sft-harm-data](https://huggingface.co/datasets/activeDap/sft-harm-data) dataset.

## Training Results

![Training Loss](loss_plot.png)

### Training Statistics

| Metric | Value |
|--------|-------|
| Total Steps | 35 |
| Final Training Loss | 2.2961 |
| Min Training Loss | 2.2961 |
| Training Runtime | 14.67 seconds |
| Samples/Second | 150.44 |

## Training Configuration

| Parameter | Value |
|-----------|-------|
| Base Model | Qwen/Qwen3-1.7B |
| Dataset | activeDap/sft-harm-data |
| Number of Epochs | 1.0 |
| Per Device Batch Size | 16 |
| Gradient Accumulation Steps | 1 |
| Total Batch Size | 64 (4 GPUs) |
| Learning Rate | 2e-05 |
| LR Scheduler | cosine |
| Warmup Ratio | 0.1 |
| Max Sequence Length | 512 |
| Optimizer | adamw_torch_fused |
| Mixed Precision | BF16 |

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "activeDap/Qwen3-1.7B_sft-harm-data"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# Format input with prompt template
prompt = "What is machine learning?\nAssistant:"
inputs = tokenizer(prompt, return_tensors="pt")

# Generate response
outputs = model.generate(**inputs, max_new_tokens=100)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)
```

## Training Framework

- **Library:** Transformers + TRL
- **Training Type:** Supervised Fine-Tuning (SFT)
- **Format:** Prompt-completion with Assistant-only loss

## Citation

If you use this model, please cite the original base model and dataset:

```bibtex
@misc{ultrafeedback2023,
      title={UltraFeedback: Boosting Language Models with High-quality Feedback},
      author={Ganqu Cui and Lifan Yuan and Ning Ding and others},
      year={2023},
      eprint={2310.01377},
      archivePrefix={arXiv}
}
```