File size: 2,846 Bytes
fb2d544 ec08c78 fb2d544 ec08c78 fb2d544 ec08c78 1505c37 ec08c78 fb2d544 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 |
---
title: LyricLoop v2.0
emoji: 🎤
colorFrom: indigo
colorTo: blue
sdk: streamlit
app_file: app.py
pinned: false
short_description: AI studio for structured lyrics, fine-tuned on Gemma-2b.
license: gemma
sdk_version: 1.53.0
---
NAME
LyricLoop LLM
---
PROJECT OBJECTIVE
LyricLoop bridges the gap between semantic LLM text and professional musical phrasing. This framework fine-tunes Google's Gemma-2b-it to generate lyrics adhering to specific structures (Verse, Chorus, Bridge) and genre-specific stylings, including Electronic, Pop, Rock, and Hip-Hop.
---
LANGUAGE / STACK
Python | PyTorch, Hugging Face (Transformers, PEFT, TRL), Streamlit
---
TECHNICAL METHODOLOGY
- Fine-Tuning: Implemented Low-Rank Adaptation (LoRA) to specialize the model in rhythmic patterns while preserving base reasoning.
- Optimization: Used 4-bit Quantization (QLoRA) via bitsandbytes to reduce the memory footprint during training.
- Instruction Tuning: Supervised Fine-Tuning (SFT) with custom templates to enforce structural and genre constraints.
---
PROJECT STRUCTURE
- app.py: main streamlit application entry point and UI logic.
- src/lyricloop/: core modular package containing engine logic:
- config.py: global constants and path management.
- data.py: prompt engineering and dataset preprocessing.
- environment.py: hardware-aware setup (MPS/CPU/CUDA).
- metrics.py: inference execution and perplexity scoring.
- viz.py: standardized plotting and visual utilities.
- notebooks/: development playground, training workflows, and EDA.
- reports/: written technical documentation and project summaries.
- assets/: visual artifacts and plots used in documentation.
- requirements.txt: dependency management for environment parity.
---
DATA & SOURCE
- Corpus: 5mm+ Song Lyrics (Genius Dataset).
- Metadata: Artist mapping via Pitchfork Reviews.
- Stack: Python, Hugging Face (Transformers, PEFT, TRL), PyTorch, and Google Colab (L4 GPU).
---
EXTERNAL RESOURCES
- Full Project Workspace (Google Drive): [Access the Notebooks & Raw Data](https://drive.google.com/drive/folders/1M5SJRaaK8OaskUgEsBupgGVN_-fQS3i4?usp=sharing)
- Training Environment: Google Colab (L4 GPU)
---
STUDIO GUIDE
- App URL: https://lxtung95-lyricloop.hf.space/
- 1. Details: Enter a song title and an Artist Aesthetic (e.g., Taylor Swift) to set the tone.
- 2. Genre: Select your target genre to adjust rhythmic density.
- 3. Compose: Use the Creativity (Temperature) slider to control experimental word choice.
- 4. Export: Download the final composition as a .txt file for your creative workflow.
---
SUPPORT
Visit my GitHub repository for the latest scripts and downloads:
https://github.com/lxntung95
--- |