---
license: apache-2.0
language:
  - en
library_name: transformers
tags:
  - smollm2
  - onnx
  - transformers.js
  - text-generation
  - fine-tuned
  - sft
  - dpo
  - lora
  - resume
  - chatbot
base_model: "HuggingFaceTB/SmolLM2-360M-Instruct"
datasets:
  - "justinthelaw/Resume-DPO-SFT-Dataset"
pipeline_tag: text-generation
---

# justinthelaw/SmolLM2-360M-Instruct_Resume-SFT-DPO

A fine-tuned version of [HuggingFaceTB/SmolLM2-360M-Instruct](https://huggingface.co/HuggingFaceTB/SmolLM2-360M-Instruct) trained with a two-stage pipeline (SFT + DPO) to answer questions about **Justin**'s professional background, skills, and experience.

## Model Description

This model is designed for browser-based inference using [transformers.js](https://huggingface.co/docs/transformers.js). It powers a personal website chatbot that can answer questions about Justin's resume, work experience, education, and skills.

### Training Pipeline

The model is trained using a two-stage approach optimized for factual memorization:

1. **SFT (Supervised Fine-Tuning)**: Primary training for factual memorization using conversation-formatted QA pairs
2. **DPO (Direct Preference Optimization)**: Refinement training to prefer accurate answers over hallucinations

### Training Details

- **Base Model**: [HuggingFaceTB/SmolLM2-360M-Instruct](https://huggingface.co/HuggingFaceTB/SmolLM2-360M-Instruct)
- **Training Dataset**: [justinthelaw/Resume-DPO-SFT-Dataset](https://huggingface.co/datasets/justinthelaw/Resume-DPO-SFT-Dataset)
- **LoRA Configuration**:
  - Rank (r): 32
  - Alpha: 64
  - Dropout: 0.05
  - Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

#### SFT Training Configuration

- Epochs: 5
- Batch Size: 8
- Learning Rate: 0.0001

#### DPO Training Configuration

- Epochs: 1
- Batch Size: 4
- Learning Rate: 5e-06
- Beta: 0.05
- Loss Type: sigmoid

## Model Formats

This repository contains multiple model formats:

| Format         | Location                     | Use Case                            |
| -------------- | ---------------------------- | ----------------------------------- |
| SafeTensors    | `/` (root)                   | Python/PyTorch inference            |
| ONNX           | `/onnx/model.onnx`           | Full precision ONNX Runtime         |
| ONNX Quantized | `/onnx/model_quantized.onnx` | Browser inference (transformers.js) |

> **Note**: If quantization fails during export due to weight distribution issues, `model_quantized.onnx` will be a copy of the fp16 model for compatibility.

## Usage

### Browser (transformers.js)

```javascript
import { pipeline } from "@huggingface/transformers";

const generator = await pipeline(
  "text-generation",
  "justinthelaw/SmolLM2-360M-Instruct_Resume-SFT-DPO",
  { dtype: "q8" } // Uses model_quantized.onnx
);

const output = await generator("What is Justin's background?", {
  max_new_tokens: 256,
  temperature: 0.7,
});
```

### Python (Transformers)

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("justinthelaw/SmolLM2-360M-Instruct_Resume-SFT-DPO")
tokenizer = AutoTokenizer.from_pretrained("justinthelaw/SmolLM2-360M-Instruct_Resume-SFT-DPO")

prompt = "What is Justin's background?"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## Intended Use

This model is intended for:

- Personal website chatbots
- Resume Q&A applications
- Demonstrating fine-tuning techniques for personalized AI assistants

## Limitations

- The model is specifically trained on Justin's resume and may not generalize to other topics
- Responses are based on training data and may not reflect real-time information
- Not suitable for general-purpose question answering

## Author

### Justin

- GitHub: [justinthelaw](https://github.com/justinthelaw)
- HuggingFace: [justinthelaw](https://huggingface.co/justinthelaw)

## License

This model is released under the Apache 2.0 license.