motd-editor-350m-ONNX

LoRA fine-tune of LiquidAI/LFM2.5-350M (run 14, data v4: v4 canon (2144 train), 6 ep, epoch picked on val by pass∧fits eval_loss/epoch=[0.2073, 0.1841, 0.1837, 0.1922, 0.2174, 0.2403] picked epoch 6 (checkpoint-3216) on val: pass∧fits 145/264, pass 178) that edits Masters of the Dungeon game definitions: bundle XML + request in, one <patch> reply out. Modified from the original: weights changed by fine-tuning, merged, and written into the onnx-community/LFM2.5-350M-ONNX graphs: onnx/model_q4.onnx (q4), onnx/model_quantized.onnx (q8), onnx/model.onnx (fp32). Licensed under the LFM Open License v1.0 (LICENSE). GGUF builds: Tim0th3us/motd-editor-350m-GGUF.

Runs in the browser with transformers.js (device: "webgpu", dtype one of the above); all keep fp32 activations, so GPUs without shader-f16 work too. motd.json holds the prompt as trained (template, system prompt, canonical input XML); constrain decoding with the schema grammar (toEbnf({ root: 'patch' }) in dungeon-core) through web-xgrammar as a logits processor.