clerc
ÉPÉE v0.3 · OPEN ASL KEYPOINT DATASET · CC BY-NC-SA 4.0

You have the model.This is the missing half.

1,200 clips of native ASL, free to download on Hugging Face. Keypoints, gloss annotations with temporal boundaries, and a frozen benchmark showing what the data alone is worth: +10 points on a signer the model has never seen, same architecture, same 24 signs.

1,200
ASL clips
×2.0 vs v0.2
6
Deaf signers
4 → 6
70.9 min
Signed duration
×1.8 vs v0.2
724
Unique glosses
×2.9 vs v0.2
128
Keypoints / frame
hands, body, face
Open
On Hugging Face
CC BY-NC-SA 4.0
CLERC · THE DATA LAYER FOR SIGN LANGUAGE AI huggingface.co/datasets/CLERC-DATA/epee
WHAT IT IS FOR · THREE USE CASES, ONE DATASET

Three products.One dataset away.

The models already exist, and you can have one running today. What no model can invent is the data underneath it. Each of these three is one model you already have, plus Épée.

+10 PTS ON A STRANGER
Small BiLSTM + clercÉpée v0.3 =Sign to word
The model is the easy part: ask Claude or ChatGPT and you have one in a minute. What it cannot invent is thousands of real signs, from different Deaf bodies, labelled frame by frame. That is Épée. Below, the same model trained twice, once on each release.
CAMERA OFF
STEP 1 · SETUP
Which hand do you sign with?
It only sets which hand goes in which slot. Nothing is recorded.
No camera needed: the replay runs real clips from a signer in neither training set.
WHAT IT READS ON A STRANGER
8 signs it reads at 70%+
overall, unseen signer59%69%
signs with 5+ examples76205
signs with 10+ examples1971
READ IT RIGHT. Listed: only signs that clear 70% recall on a signer the model never saw, with the precision to match. 24 signs is the frozen benchmark, not the corpus ceiling: depth makes a wider vocabulary trainable, it does not mean it is trained. One sign, one word is the first step of translation, not translation: sentences need grammar on top, which is what the parallel grid trains.
TRAINED ON ÉPÉE v0.2
·
replay a clip to compare
TRAINED ON ÉPÉE v0.3
·
replay a clip to compare
ALL FIGURES COMPUTED FROM BOTH PUBLISHED RELEASES huggingface.co/datasets/CLERC-DATA/epee
THE EVOLUTION BEHIND IT · PROTOCOL FROZEN SINCE v0.2

Twice the data.Ten points on a stranger.

Same small BiLSTM, same 24 signs, same held-out Deaf signers, same seeds. The only thing that changed is the dataset.

ACCURACY ON A SIGNER OUTSIDE THE DATASET · BY # TRAINING SIGNERS

mean over all combinations · 8 seeds · chance 4.2%
[ VIEW AS TABLE ]

ACCURACY · BY AMOUNT OF DATA

same unseen signer, growing pool
[ VIEW AS TABLE ]
59%69%

On a signer never seen. Second unseen signer: 54% → 58%.

433 pts

Gap to the known-signer ceiling (73%), with 1 vs 6 training signers.

±15±5

Run-to-run spread, 1 vs 6 signers. More bodies make the result dependable.

READ IT HONESTLY. 24-gloss vocabulary frozen from v0.2 · unseen means zero shared clips (prompts are shared by protocol) · the five grid signers alone reach 70.5%, level with six at 69.4%: FOXTROT adds a second register and vocabulary, not points on this conversational test · a bigger model does not close the one-vs-six gap. It is a data property.
SAME MODEL · SAME TEST · ONLY THE DATA CHANGED huggingface.co/datasets/CLERC-DATA/epee
THREE RELEASES · ELEVEN WEEKS · SINCE JUNE 2026

Doubled twice.Same protocol throughout.

Same repo, same format, same pseudonymized signers since v0.1. v0.3 is main; every version stays pinned under its git tag and Zenodo DOI. Pick a release.

TAGS: v0.1 · v0.2 · v0.3 (MAIN) DOI 10.5281/zenodo.22081248
QUESTIONS · BEFORE YOU DOWNLOAD

Asked often.Answered once.

What is in the CLERC Épée sign language dataset?
1,200 ASL clips from 6 native Deaf signers: 70.9 minutes of signing, 128 motion keypoints per frame (hands, body, face, head), and ASL gloss annotations with temporal segmentation. A 201-phrase parallel grid lets you compare the same phrase across signers directly. The public release ships extracted keypoints and annotations only; raw video stays under commercial license.
How do I download the ASL dataset?
The Épée release is open on Hugging Face (CLERC-DATA/epee) under CC BY-NC-SA 4.0: plain .npy and .json files, loadable in a few lines of Python with huggingface_hub. Full corpus access and commercial licensing: florian@clerc.io.
Why does multi-signer ASL training data matter?
The benchmark holds test signers entirely outside the training set: accuracy on a brand-new signer climbs from 29% with 1 training signer to 69% with 6, with the same model and hyperparameters. A model trained on one signer has learned one person, not a language, which is why data, not architecture, is the bottleneck for sign language AI.
What changed between Épée v0.2 and v0.3?
The corpus doubled: 600 to 1,200 clips, 4 to 6 Deaf signers, 251 to 724 unique glosses, and the parallel grid grew from 150 to 201 phrases. On the frozen benchmark, accuracy on an unseen signer went from 59% to 69%, and the signs it reads reliably went from 5 to 8. Previous versions stay pinned under git tags and Zenodo DOIs.
Can I train a commercial model on this data?
The public release is non-commercial (CC BY-NC-SA 4.0): research, education and prototyping are covered. Training commercial models or foundation models requires a commercial license, which also unlocks the full multi-signer corpus beyond this release.
OPEN RELEASE · CC BY-NC-SA 4.0

Free to use.One line to load.

Keypoints and gloss annotations, numpy-ready. No raw video ships: signers stay pseudonymized, and every number on this page is reproducible from the download.

QUICK START · PYTHONmetadata.csv · keypoints/*.npy
from huggingface_hub import snapshot_download import numpy as np, pandas as pd root = snapshot_download(repo_id="CLERC-DATA/epee", repo_type="dataset") meta = pd.read_csv(f"{root}/metadata.csv") # five renderings of the same phrase, ready to compare phrase = meta[(meta.parallel == "yes") & (meta.phrase_id == 7)] clips = {r.signer_id: np.load(f"{root}/keypoints/{r.clip_id}.npy") for r in phrase.itertuples()} # each (n_frames, 128, 3)
Research and evaluation are free under CC BY-NC-SA 4.0, with attribution and the DOI. The video behind these keypoints and the full CLERC corpus (14,400+ videos, 16 signers, growing weekly) are licensed commercially: florian@clerc.io
CLERC · SIGN LANGUAGE DATA INFRASTRUCTURE CLERC.IO