| --- |
| title: Datasent |
| emoji: 🟪 |
| colorFrom: indigo |
| colorTo: blue |
| sdk: static |
| pinned: true |
| --- |
| |
| <div align="center"> |
| <img src="https://huggingface.co/spaces/datas3nt/README/resolve/main/datasent_banner.png" alt="Datasent - Polygen: polynomial tokenization and compression for ML data" width="920" style="width:100%; max-width:920px;" /> |
| </div> |
|
|
| Datasent builds **Polygen**, an SDK that tokenizes numeric, image, text, audio, |
| and video data by fitting a compact basis to each window, with a lossless |
| `x = f_x + r` decomposition: the fitted prediction `f_x` plus the residual `r`. |
| The representation is small, structured, and decodes exactly. |
|
|
| ## Polygen for Qwen2-VL |
|
|
| Our flagship demo applies Polygen to a vision-language model: it compresses the |
| visual tokens Qwen2-VL feeds to the LLM, and a small LoRA adapter teaches the |
| model to read the compressed representation. On ScienceQA the recipe holds or |
| improves accuracy while shrinking the visual sequence: |
|
|
| - **+8.4pp accuracy** over the un-spliced baseline at the accuracy tier (T=3) |
| - At the compression tier (T=2, +7.4pp accuracy): **~49% fewer visual tokens** and **~3.7x (nearly 4x) more requests per GPU** in cached serving |
| - **~5x smaller** stored visual representation |
|
|
| The vision encoder and LLM weights are otherwise untouched; the compression sits |
| between the merger and the LLM. |
|
|
| ## Try it |
|
|
| - **[Polygen for Qwen2-VL](https://huggingface.co/spaces/datas3nt/qwen2vl-polygen)** -- upload an image, compare answers with and without compression, scored per image |
| - **[Polygen RAG demo](https://huggingface.co/spaces/datas3nt/qwen2vl-polygen-rag)** -- retrieval over a compressed visual-token corpus |
| - **[Polygen tokenization demo](https://huggingface.co/spaces/datas3nt/polygen-demo)** -- the core tokenizer |
|
|
| ## Models |
|
|
| - **[qwen2vl-polygen-7b-lora-r16-1000](https://huggingface.co/datas3nt/qwen2vl-polygen-7b-lora-r16-1000)** -- 7B LoRA adapter |
| - **[qwen2vl-polygen-lora-r16-1000](https://huggingface.co/datas3nt/qwen2vl-polygen-lora-r16-1000)** -- 2B LoRA adapter |
|
|
| Learn more at [datasent.com](https://www.datasent.com). |
