Datasets:

---

CLERC-DATA

/

epee

Like 5

Follow

CLERC 2

v0.4 is live: 2,400 ASL clips, 8 Deaf signers, 300 clips each, a 302-phrase parallel grid over 5 signers. Twice the clips of v0.3, two more signers, 1,501 unique glosses, 179.9 minutes of signing.

Before the curve, a correction.

v0.1 to v0.3 shipped with crossed hand slots

The 0-20 / 21-41 keypoint blocks were supposed to hold the same hand on every clip. They did not. Measured on the published v0.3 files, 61.7% of the clips whose convention can be decided had the two hands swapped, and the rate differed by signer. The swap correction existed in our build, a bug kept it from firing, and nothing checked the built files.

v0.4 anchors every clip to the pose wrists (0-20 is the hand at the signer's left wrist, 21-41 at the right wrist) and re-checks the built files independently before release: 0 crossed out of 2,395 decidable clips. A new hand_slots column says verified, or unverified for the 5 clips with too little hand evidence to decide.

If you trained on v0.1 to v0.3, re-download. A model trained on mixed slots learns "which slot holds the active hand" as a signer cue, and pays for it on every new signer.

What the fix is worth, and what the data is worth

Same small BiLSTM, same 24 glosses, same test segments and same 8 seeds as the v0.2 and v0.3 measurements, tested on two Deaf signers with no clips in the release. We separated the two effects instead of reporting one jump:

On Signer A: +6.3 points from the correction alone (same clips), +5.7 from the new data. The 12 points are not a data effect and we will not quote them as one.

We also held volume fixed and changed only the signer mix: at equal volume, eight signers beat six by 2.3 points on Signer A, within run-to-run spread. This run does not separate diversity from volume, so we do not claim it does.

The curve

Trained on 1 signer, then 2, up to 8, tested each time on a signer held entirely outside the training set:

macro-F1 over the same range: 0.28 to 0.74. A signer the model has already seen scores 82.5%. At eight training signers, the stranger is at 81.4%.

This curve is not comparable to the one in #4. That one was trained and tested on crossed slots.

Full method, exact numbers and caveats: BENCHMARK.md.

What else is in v0.4

• Two new Deaf signers: GOLF (fluent ASL, L2) and HOTEL (native, ASL L1), each on an own prompt set of declarative sentences. Signer pool 6 to 8, 4 female / 4 male.
• Parallel grid widened 201 to 302 phrases across ALPHA to ECHO. Pair renderings through phrase_id.
• FOXTROT 200 to 300 clips on his own register. Three own registers in total (FOXTROT, GOLF, HOTEL), disjoint from the grid and from each other. Filter on parallel.
• Clips renumbered clerc_v04_0001 to clerc_v04_2400. v0.1, v0.2 and v0.3 stay reachable as tags.
• One v0.3 recording withdrawn for privacy (a fill-in prompt answered with personal information). It is not in v0.4 and has been removed from the v0.3 tag.
• No clip and no new phrase of the sequestered Épée 200 exam is published.

What we did not do

Dominant hand is not normalized. Slots are fixed to the body, not to handedness: left-dominant signers are not mirrored. That modelling choice is left to you.

This is still a pilot. 2,400 clips and 8 signers is a baseline, not a production corpus. The benchmark measures generalization across signer bodies on a 24-gloss subset, not across topics.

Skeletons are body data. On v0.3, a plain logistic regression separated the six signers at 94% from nine limb-ratio features. Pseudonymization protects names, not bodies.

Live demo, dictionary and replay: clerc.io/data

CC BY-NC-SA 4.0 · DOI 10.5281/zenodo.20268565 (resolves to the latest version, a version DOI for v0.4 will follow) · commercial licensing: florian@clerc.io

If you trained on an earlier release and your numbers move after re-downloading, tell us here. That is exactly the report we want.