emberian/h-leaf-e4-decay10-onnx

ONNX export of a Falcon-H1-Tiny checkpoint from the h project (https://github.com/emberian/h), produced by site/export/export_onnx.py for transformers.js 4.x. Source checkpoint: trunk-wsd-lr1e-4-seed0 final checkpoint (4 epochs = 1.498B corpus-v1 tokens, LR 1e-4 constant) cooled over 37.4M tokens (leaf-e4-decay10); validation 3.136 on 512 seq; see github.com/emberian/h research/results/tpu-h1jax-trunk.

Files: onnx/model.onnx (fp32) and onnx/model_quantized.onnx (8-bit MatMulNBits + 8-bit block-quantized embedding; near-lossless for this model). The 4-bit variant, if present, is provided for comparison only: 4-bit round-to-nearest degrades this 90M model badly.

Derived from tiiuae/Falcon-H1-Tiny-90M-Base under the Falcon-LLM License; the license, acceptable-use policy, and redistribution notice are included in this repository.

const model = await AutoModelForCausalLM.from_pretrained("emberian/h-leaf-e4-decay10-onnx", { dtype: "q8", device: "webgpu" });

Export manifest: {"checkpoint": "/Users/ember/dev/h/artifacts/checkpoints/leaf-e4-decay10", "versions": {"python": "3.12.13", "platform": "macOS-26.6.1-arm64-arm-64bit", "torch": "2.13.0", "transformers": "4.57.6", "onnx": "1.22.0", "onnxruntime": "1.29.0", "onnxslim": "0.1.96", "onnx-ir": "1.0.0", "numpy": "2.5.2", "safetensors": "0.8.0", "tokenizers": "0.22.2"}, "opset": 18, "command": "site/export/export_onnx.py artifacts/checkpoints/leaf-e4-decay10 site/models/h-leaf-e4-decay10", "parameters": 91131072, "export_s": 40.1, "slim_s": 22.5, "fp32_ops": ["Abs", "Add", "And", "Cast", "Clip", "Concat", "Cos", "CumSum", "Div", "Equal", "Exp", "Expand", "Gather", "Greater", "GreaterOrEqual", "LessOrEqual", "Log", "MatMul", "Mul", "Neg", "Pow", "Range", "ReduceMean", "Relu", "Reshape", "Shape", "Sigmoid", "Sin",