openlocalagent-10m-matrix · WebGPU bundle
The browser bundle of danelcsb/openlocalagent-10m-matrix:
a 10.5M-parameter tool-calling agent trained from scratch on open data, exported to ONNX (fp16)
for onnxruntime-web with the WebGPU execution provider. It is the model behind the OpenLocalAgent
workbench demo, which runs entirely in the browser.
| file | purpose |
|---|---|
model.fp16.onnx |
full-sequence logits graph |
cached/prefill.fp16.onnx, cached/decode.fp16.onnx |
prefill-then-KV-cached decode graphs |
cached/meta.json, cached/provenance.json, cached/single-decode.json |
graph contract and pinned hashes |
tokenizer.json |
the 16,384-token ByteLevel BPE the model was trained with (sha256 2f49a644…) |
meta.json |
markers, tool catalog the demo starts from, model parameters |
bundle-manifest.json |
sha256 of every artifact, checkpoint identity (posttrain-la-10m-matrix, step 1199) |
Prompt contract
openai_full_catalog_v1: the tool catalog is rendered into the prompt as canonical JSON inside
<|tool_catalog|>…</|tool_catalog|>, followed by <|user|> request <|assistant|>, and the model
answers with <tool_call>{"arguments":{…},"name":"…"}</tool_call> or plain text. Tool responses are
fed back as <|tool|><tool_response>…</tool_response>. The context is 2,048 tokens; keep the
catalog small (a few tools), which is what the model was trained on.
Use
Point the demo at this revision:
workbench.html?bundle=https://huggingface.co/danelcsb/openlocalagent-10m-matrix-webgpu/resolve/main/
Source, training recipe, evaluation and the workbench: the OpenLocalAgent repository.
Provenance
Pretrain on a five-source open matrix (FineWeb-Edu, C4, Wikipedia, Cosmopedia, FineMath; 4.4B tokens), midtrain on non-benchmark agent data, posttrain on the train splits of the scored benchmarks. The ten-benchmark receipt for this checkpoint: sound-tool-calling mean 19.1, legacy ten-suite mean 17.6 (three seeds: 17.6 / 16.5 / 16.8).