MiniCheck-Flan-T5-Large (ONNX)
An ONNX export of
MiniCheck-Flan-T5-Large
(770M, FLAN-T5-large fine-tuned for grounded fact-checking), for in-process
inference without a Python/PyTorch runtime.
Why this exists
Upstream MiniCheck publishes only the PyTorch checkpoint - no ONNX build existed
for this 770M encoder-decoder verifier. This is that build, so anyone who wants
to experiment with MiniCheck-Flan-T5-Large can, without standing up a PyTorch
runtime or writing a custom export pathway.
It reflects a Familiar Tools belief: a specialized, right-sized model that runs
efficiently and in-process beats reaching for a large, general, resource-hungry
one. Exporting a focused model to ONNX is part of that - it makes the model
cheap to run, easy to embed, and light on dependencies. Custom, deliberately
engineered solutions tend to be more efficient and more resource-aware than
general-purpose defaults.
Files
Exported with optimum-cli export onnx --task text2text-generation --opset 14 --dtype fp32.
| File | Notes |
|---|
encoder_model.onnx (~1.3 GB) | Encoder stack. Inputs: input_ids, attention_mask. |
decoder_model.onnx (~1.8 GB) | Decoder stack. Inputs: input_ids, encoder_hidden_states, encoder_attention_mask. No KV-cache (single decode step). |
tokenizer.json | HF fast tokenizer (loads with the Rust tokenizers crate). |
spiece.model | SentencePiece model. |
config.json, generation_config.json, tokenizer_config.json, special_tokens_map.json | T5 config + tokenizer metadata. |
MiniCheck is used as a single-step seq2seq verifier: the encoder reads the
(document, claim) prompt, the decoder takes one step, and the probability mass
on the "Yes" vs "No" vocabulary tokens gives the support score.
License and attribution
Released under the MIT License, matching upstream.
- MiniCheck: Tang et al., MiniCheck: Efficient Fact-Checking of LLMs on
Grounding Documents, EMNLP 2024
(arXiv:2404.10774).
- Original weights:
lytang/MiniCheck-Flan-T5-Large.
- Base model:
google/flan-t5-large.
This repo redistributes a derivative (ONNX export) of the above under the same
MIT terms. Weights were not retrained or modified; only the inference graph was
re-expressed.
