---
license: apache-2.0
tags:
  - text-to-json
  - music-preferences
  - fine-tuned
  - bart-large
---

# Model Card for afsagag/bart-spotify-preferences

This is a fine-tuned BART-large model for converting music prompts to Spotify feature preferences (e.g., Energy, Valence, Release Year). Fine-tuned on a ~1k sample dataset in Kaggle.

## Training Details
- **Dataset**: PromptsToSpotifyFeatures-v2 (~1k samples)
- **Framework**: Hugging Face Transformers
- **Hyperparameters**:
  - Learning rate: 5e-05
  - Epochs: 7
  - Batch size: 4 (with gradient accumulation steps=4)
  - FP16: True
- **Metrics**: MAE, RMSE, per-feature correlation

## Usage
```python
from transformers import BartForConditionalGeneration, BartTokenizer
model = BartForConditionalGeneration.from_pretrained("afsagag/bart-spotify-preferences")
tokenizer = BartTokenizer.from_pretrained("afsagag/bart-spotify-preferences")
prompt = "music for a supervillain"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=256, num_beams=1, length_penalty=0.6, no_repeat_ngram_size=2, early_stopping=True)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## Notes
- Fine-tuned without LoRA for full model weights.
- Outputs JSON-like dictionaries; may require post-processing for malformed JSON.
- Trained on Kaggle T4 GPU with ~16GB VRAM.