---
license: apache-2.0
base_model: Qwen/Qwen2.5-3B-Instruct
tags:
- peft
- lora
- value-stream-mapping
- manufacturing
- qwen2.5
library_name: peft
---

# VSM-LLM-3B-Fast

LoRA adapter for Qwen2.5-3B-Instruct, fine-tuned for Value Stream Mapping (VSM) generation.

## Model Details

- **Base Model**: Qwen/Qwen2.5-3B-Instruct
- **Training Method**: LoRA (r=8, alpha=16)
- **Task**: Generate structured JSON for VSM plots
- **Training Data**: 20 synthetic VSM examples
- **Trainable Parameters**: 14.9M (0.48% of base model)

## Usage

### With Transformers + PEFT

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-3B-Instruct")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-3B-Instruct")

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "nishit1945/VSM-LLM-3B-Fast")

# Generate
messages = [
    {"role": "system", "content": "You are a VSM expert."},
    {"role": "user", "content": "Generate VSM JSON..."}
]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt")
outputs = model.generate(inputs, max_new_tokens=400, temperature=0.1)
```

### HuggingFace Inference API

```bash
curl https://router.huggingface.co/hf-inference/models/nishit1945/VSM-LLM-3B-Fast \
  -H "Authorization: Bearer YOUR_HF_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"inputs": "VSM prompt...", "parameters": {"max_new_tokens": 400, "temperature": 0.1}}'
```

## Training Details

- Epochs: 3
- Batch Size: 1 (gradient accumulation: 16)
- Learning Rate: 2e-4
- Optimizer: paged_adamw_8bit
- Quantization: 4-bit (training only)

## Performance

- **Inference Time**: 2-4 seconds on CPU
- **Output Format**: Structured JSON with processes, flows, metrics, coordinates

## License

Apache 2.0