Spaces:
Sleeping
Sleeping
| title: TensorPEFT | |
| emoji: 🧊 | |
| colorFrom: purple | |
| colorTo: blue | |
| sdk: gradio | |
| sdk_version: 6.31.0 | |
| app_file: app.py | |
| pinned: false | |
| license: apache-2.0 | |
| short_description: Tensor-pursuit fine-tuning vs LoRA at tiny budgets | |
| tags: | |
| - peft | |
| - fine-tuning | |
| - tensor-decomposition | |
| - lora-alternative | |
| # TensorPEFT live demo | |
| Live race between **TensorPEFT** (tensor gradient-pursuit adapters) and matched-budget PEFT baselines | |
| (prompt tuning, LoRA, (IA)³, LayerNorm tuning): GPT-2 124M is fine-tuned on Bengali news with the same number of | |
| trainable parameters for every method, and validation cross-entropy decides. | |
| - Library: [Angshul/TensorPEFT](https://huggingface.co/Angshul/TensorPEFT) — `pip install git+https://huggingface.co/Angshul/TensorPEFT` | |
| - Data: Bengali news articles from [soham96/Bengali_news_classifier](https://github.com/soham96/Bengali_news_classifier), | |
| downloaded at start-up (not redistributed), tokenised with the GPT-2 tokenizer; 3M training tokens, held-out articles for validation. | |
| - Base model: [openai-community/gpt2](https://huggingface.co/openai-community/gpt2). | |
| Runs on ZeroGPU. Benchmark numbers in the *Benchmark results* tab were measured on CPU (fp32, 300 steps). | |