Modelpublic

Meddies/meddies-title-v1-onnx

sourceHugging Faceotherupdated 5d agoView on Hugging Face

0likes6downloads

Model Card

DRAFT — Meddies Title v1 ONNX (for review, not yet public)

This card is a draft. Do not treat numbers below as final until review signs off.

CPU-serving build of the Meddies Title v1 model: the merged epoch-1 weights exported to ONNX FP16 with Liquid's lfm2-export. For model weights, training lineage, and the full evaluation story, see Meddies/meddies-title-v1.

Files

File	What it is
`onnx/model_fp16.onnx` (+ `model_fp16.onnx_data`, 454 MB total)	FP16 ONNX graph + external weights
`tokenizer.json`, `tokenizer_config.json`, `chat_template.jinja`	Tokenizer pinned to the training setup
`config.json`, `generation_config.json`	Model config carried for loaders

Serving

CLI (LiquidONNX):

bash

lfm2-infer --model onnx/model_fp16.onnx --prompt "YOUR SESSION TEXT HERE" --cpu

Python (verified pattern — this exact loop generated the parity set below):

python

from liquidonnx.session import ONNXTextModel  # https://github.com/Liquid4All/onnx-export

SYSTEM = "Write one clear, concise, memorable session title in the query's primary language. Make it specific enough to recognize months later. Use the subject, distinguishing task, and only a necessary qualifier. Name what the user wants without answering the query. Use only stated facts. Drop greetings, politeness, and irrelevant background. Do not repeat the full query. Output only the title."

model = ONNXTextModel("onnx/model_fp16.onnx", force_cpu=True)
model.load()
title = model.generate(
    [{"role": "system", "content": SYSTEM},
     {"role": "user", "content": "YOUR SESSION TEXT HERE"}],
    max_new_tokens=128,
    stream=False,
)

Parity (measured)

85 validation queries (5/language), same frozen prompt, greedy, FP16 ONNX vs FP32 adapter: 52/85 token-identical generations, 11/85 exact vs reference (FP32: 10/85), zero blanks. All divergences are same-quality paraphrases — the export is faithful.

License

Derived from LiquidAI/LFM2.5-230M, so redistribution follows the LFM Open License v1.0: free commercial use below US$10M annual revenue with attribution and change notices; a separate commercial license is required above that threshold. See https://www.liquid.ai/lfm-license for the full terms.