Add browser ONNX export with INT8 embeddings and usage guide

onnx_browser/README.md

@@ -0,0 +1,38 @@
+# Bekko System One v0 17M — browser ONNX
+
+This directory contains an ONNX conversion for the dedicated **Bekko System One browser runtime**. It supports Choice, Noul, and Score typed decisions. The companion Node.js runtime can also load this directory.
+
+This is not a Sentence Transformers ONNX backend export or a general-purpose sentence embedding model. Use the model repository's root files and `inference_v0.py` for Python / Sentence Transformers inference.
+
+## Files and precision
+
+| File | Purpose |
+| --- | --- |
+| `model.onnx` | Shared-prefix encoder and three prediction heads |
+| `manifest.json` | Model filename, hashes, task order, token budgets, quantization policy, and source revision |
+| `tokenizer.json` | Tokenizer vocabulary and processing rules |
+| `tokenizer_config.json` | Special tokens used by the dedicated runtime |
+
+Only the token embedding table is stored as row-wise symmetric INT8. Transformer blocks and prediction heads remain FP32; this is not fully INT8 inference. The tokenizer is unchanged. The ONNX file is 29.0 MB (decimal), excluding tokenizer and runtime downloads.
+
+## Loading
+
+Use the [Bekko browser runtime and usage guide](https://github.com/hotchpotch/bekko-system-one/tree/main/browser). Set the model base URL to:
+
+```text
+https://huggingface.co/hotchpotch/bekko-system-one-v0-17m/resolve/<COMMIT_REVISION>/onnx_browser/
+```
+
+Replace `<COMMIT_REVISION>` with a commit containing this directory and pin that same revision for all files. The runtime reads `manifest.json`, both tokenizer files, and the model named in the manifest. It handles typed-decision input rendering, candidate grouping, and output probabilities. Loading `model.onnx` alone does not provide those application-level operations.
+
+Browser execution uses ONNX Runtime Web. Available execution providers, memory use, and practical input sizes depend on the browser and device. Download size is not peak runtime memory; larger models may exceed a device's memory budget. Node.js usage is documented in the same guide.
+
+## Conversion and verification
+
+- [FP32 ONNX export script](https://github.com/hotchpotch/bekko-system-one/blob/main/browser/scripts/export_onnx.py)
+- [Embedding INT8 conversion script](https://github.com/hotchpotch/bekko-system-one/blob/main/browser/scripts/quantize_onnx.py) (`--embedding-only`)
+- [ONNX verification script](https://github.com/hotchpotch/bekko-system-one/blob/main/browser/scripts/verify-export.js)
+
+The INT8 output is distributed here as `model.onnx`; the manifest records its quantization policy. FP32 reference files and parity fixtures are retained separately for verification.
+
+Source model revision: `2c3f04e55e8d0764620ee46fc86be33e408d1ebd`. The export checkpoint weights and tokenizer were checked against this revision. On 5 synthetic fixtures, Node.js / ONNX Runtime verification measured a maximum FP32 logit error of 3.3378601e-06 against the Python reference, and a maximum absolute probability difference of 0.0043458041 between FP32 and embedding-INT8 ONNX. These checks are not a full quality evaluation or a guarantee of browser/device compatibility.

onnx_browser/manifest.json

@@ -0,0 +1,27 @@
+{
+  "format_version": 1,
+  "tasks": [
+    "choice",
+    "noul",
+    "score"
+  ],
+  "query_length": 4096,
+  "document_length": 2048,
+  "cls_token_id": 50281,
+  "sep_token_id": 50282,
+  "pad_token_id": 50283,
+  "model_sha256": "1a27d9ce0f44ddb8e3247aec245e3c05e2425996d5ff0def114a093c58c464de",
+  "heads_sha256": "03aaa7864b6fd5c08c5ca57f233b65f907226ad43579c91f0998f5eaf51f7aac",
+  "weights_sha256": "e4c2587853fa9073b022c18801857733abf2950bb765fe37e434b6baaae18ac7",
+  "model_file": "model.onnx",
+  "model_bytes": 29023607,
+  "source_model_sha256": "b3b55bba127f6d3eca5c3f4762e04b77b1bef301e3242132a58fca5e93813b52",
+  "quantization": {
+    "embedding": "rowwise_symmetric_int8",
+    "blocks": "fp32",
+    "heads": "fp32",
+    "integer_matmuls": 0
+  },
+  "source_repository": "hotchpotch/bekko-system-one-v0-17m",
+  "source_revision": "2c3f04e55e8d0764620ee46fc86be33e408d1ebd"
+}

onnx_browser/model.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a27d9ce0f44ddb8e3247aec245e3c05e2425996d5ff0def114a093c58c464de
+size 29023607

onnx_browser/tokenizer_config.json

@@ -0,0 +1,7 @@
+{
+  "cls_token": "[CLS]",
+  "sep_token": "[SEP]",
+  "pad_token": "[PAD]",
+  "unk_token": "[UNK]",
+  "mask_token": "[MASK]"
+}