Add Datasent organization card
Browse files
README.md
CHANGED
|
@@ -1,10 +1,48 @@
|
|
| 1 |
---
|
| 2 |
-
title:
|
| 3 |
-
emoji:
|
| 4 |
-
colorFrom:
|
| 5 |
-
colorTo:
|
| 6 |
sdk: static
|
| 7 |
pinned: false
|
| 8 |
---
|
| 9 |
|
| 10 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
+
title: Datasent
|
| 3 |
+
emoji: 🟪
|
| 4 |
+
colorFrom: indigo
|
| 5 |
+
colorTo: blue
|
| 6 |
sdk: static
|
| 7 |
pinned: false
|
| 8 |
---
|
| 9 |
|
| 10 |
+
<div align="center">
|
| 11 |
+
|
| 12 |
+
# Datasent
|
| 13 |
+
|
| 14 |
+
### Polygen: polynomial tokenization and compression for ML data
|
| 15 |
+
|
| 16 |
+
</div>
|
| 17 |
+
|
| 18 |
+
Datasent builds **Polygen**, an SDK that tokenizes numeric, image, text, audio,
|
| 19 |
+
and video data by fitting a compact basis to each window, with a lossless
|
| 20 |
+
`x = f_x + r` decomposition: the fitted prediction `f_x` plus the residual `r`.
|
| 21 |
+
The representation is small, structured, and decodes exactly.
|
| 22 |
+
|
| 23 |
+
## Polygen for Qwen2-VL
|
| 24 |
+
|
| 25 |
+
Our flagship demo applies Polygen to a vision-language model: it compresses the
|
| 26 |
+
visual tokens Qwen2-VL feeds to the LLM, and a small LoRA adapter teaches the
|
| 27 |
+
model to read the compressed representation. On ScienceQA the recipe holds or
|
| 28 |
+
improves accuracy while shrinking the visual sequence:
|
| 29 |
+
|
| 30 |
+
- **+8.4pp accuracy** over the un-spliced baseline at the accuracy tier (T=3, n=200)
|
| 31 |
+
- At the compression tier (T=2, +7.4pp accuracy): **~49% fewer visual tokens** and **~3x more requests per GPU** in cached serving
|
| 32 |
+
- **~5x smaller** stored visual representation
|
| 33 |
+
|
| 34 |
+
The vision encoder and LLM weights are otherwise untouched; the compression sits
|
| 35 |
+
between the merger and the LLM.
|
| 36 |
+
|
| 37 |
+
## Try it
|
| 38 |
+
|
| 39 |
+
- **[Polygen for Qwen2-VL](https://huggingface.co/spaces/datas3nt/qwen2vl-polygen)** -- upload an image, compare answers with and without compression, scored per image
|
| 40 |
+
- **[Polygen RAG demo](https://huggingface.co/spaces/datas3nt/qwen2vl-polygen-rag)** -- retrieval over a compressed visual-token corpus
|
| 41 |
+
- **[Polygen tokenization demo](https://huggingface.co/spaces/datas3nt/polygen-demo)** -- the core tokenizer
|
| 42 |
+
|
| 43 |
+
## Models
|
| 44 |
+
|
| 45 |
+
- **[qwen2vl-polygen-7b-lora-r16-1000](https://huggingface.co/datas3nt/qwen2vl-polygen-7b-lora-r16-1000)** -- 7B LoRA adapter
|
| 46 |
+
- **[qwen2vl-polygen-lora-r16-1000](https://huggingface.co/datas3nt/qwen2vl-polygen-lora-r16-1000)** -- 2B LoRA adapter
|
| 47 |
+
|
| 48 |
+
Learn more at [datasent.com](https://www.datasent.com).
|