Spaces:
Sleeping
Sleeping
Commit History
15 days ago
-
fix: bound a pair at 512 tokens, as meaningbert.scorer does; DeBERTa-v3 reports no limit
Verified signature4d433e3 -
fix: an exact match is the same sentence at the same position, not a reference repeated in another pair
Verified signature8a9637d -
fix: apply the checkpoint's output head, so large scores on 0-100 like v1; v1 untouched
Verified signature5f31a78 -
fix: the widget and the usage example call compute with references and predictions
Verified signatureac07815 -
add device handling: cuda when available, cpu otherwise, explicit device on compute
Verified signature5808a3c
23 days ago
-
load the v2 checkpoints from subfolders of davebulaval/MeaningBERT
Both v2 checkpoints live in one repository, as subfolders, so the article, the card and the numbers stay in one place. The root of that repository is still the v1 model, untouched: anything already calling evaluate.load without a name keeps the scores it had. Names follow the convention the field already reads: large and base, not best and fast. best ages badly, since the day a better model ships the name is either wrong or has to be overwritten under the users who depend on it. base is not published yet and raises. v1 is reachable by name for reproducibility.6419cbb -
point the metric at MeaningBERT-v2
The default variant now loads davebulaval/MeaningBERT-v2: deberta-v3-large trained on the v2 corpus. On 1536 human-annotated test pairs it correlates at 0.704 against 0.323 for the v1 model, and scores 97.1 % of identical pairs above 95 against 0 %. 'fastest' still raises: that checkpoint ships later.6a040ee -
fix: rebuild the widget on current Gradio, and name the two variants
The Space died at startup: evaluate.utils.launch_gradio_widget still builds its interface with gr.inputs.Dataframe, removed in Gradio 4, while this Space declares sdk_version 5.9.1. AttributeError: module 'gradio' has no attribute 'inputs' Pinning gradio<4 would fix it too, on an unmaintained major version that will break again. Forty lines of current Gradio is the cheaper of the two, and the widget now reports an unusable row instead of raising a stack trace on a public page. evaluate.load also takes a variant as its second argument. Only 'best' is published today; 'fastest' ships with MeaningBERT v2 and RAISES until then. A silent fallback to the other checkpoint would let someone publish scores from a model they did not ask for.050aed1
-
Update README.md
Verified signature6689ab9
-
remove device handling
davebulaval495f38b
-
add torch.no_grad()
davebulavalba092fc
-
fix deprecated keyed elements
davebulaval1541189 -
fix inversion of reference and prediction
davebulaval364be12 -
uniformization of interface and add .to for tokenizer output
davebulaval7217d6a -
Add device and remove dead verbose argument.
Verified signature1982c24
-
add fix for matching elements
davebulaval46c63a3 -
add fix for matching elements
davebulaval008aa62 -
Update meaningbert.py
Verified signature35bc035 -
Update meaningbert.py
Verified signature09175db -
Update meaningbert.py
Verified signature18e88dc -
Update requirements.txt
Verified signatureb4ffb97 -
Update README.md
Verified signature14a7306 -
Update requirements.txt
Verified signaturee1ba24b -
Update requirements.txt
Verified signature30ce2a1 -
Update requirements.txt
Verified signaturec456f0b
-
improve processing and doc
davebulaval4e0f879
-
try fix app problem
davebulaval04f1736 -
try fix app problem
davebulavald13980c -
try fix app problem
davebulavalc466d08 -
try fix app problem
davebulavalb50fe98 -
add missing dependencies
davebulaval13e0fc1 -
add missing dependencies
davebulaval7c70256 -
fix typos
davebulavaldf39cd4 -
initial metrics codebase
davebulaval9db8db7
-
initial commit
7cd4a72