---
license: apache-2.0
base_model: google/bert_uncased_L-4_H-256_A-4
library_name: onnx
pipeline_tag: feature-extraction
tags:
- sentence-transformers
- embeddings
- scientific-papers
- distillation
- onnx
- int8
- wasm
- research-library
datasets:
- PeytonT/1m_papers_text
- PeytonT/paper_universe_interactive
language:
- en
---

# M1-Lite Paper Embedding Model ONNX

This is a compact student encoder distilled from `PeytonT/1m-paper-embedding-model` over the 1M paper corpus.
It predicts the same 768-dimensional, L2-normalized M1 embedding space used by the Research Library paper universe.

## Files

| Path | Description |
| --- | --- |
| `onnx/model.onnx` | Float ONNX student encoder. |
| `onnx/model.int8.onnx` | Dynamically quantized int8 ONNX model for browser/WASM inference. |
| `tokenizer/` | Student tokenizer files. |
| `manifest.json` | Export metadata consumed by the static viewer. |

## Training

- teacher target: `exports/huggingface/paper_universe_interactive_v1/semantic_m1/papers_all.emb.i8`
- interactive level: `exports/huggingface/paper_universe_interactive_v1/interactive/papers_all.json`
- rows: `1,000,000`
- final eval cosine vs M1 targets: `0.7351`
- final eval MSE vs M1 targets: `0.000690`
- int8 ONNX size: `33.7 MB`

## Embedding Format

- output name: `embedding`
- shape: `[batch, 768]`
- pooling: attention-mask mean pooling, projection to 768 dimensions, L2 normalization
- intended max sequence length: `128`

The output is compatible with the existing `semantic_m1/*.emb.i8` paper-vector files.