CLERC-DATA/epee-v01

CLERC Épée v0.1

The first AI-grade sign language data layer.

A pilot release of structured, multi-signer ASL data designed for AI training, research benchmarking, and inter-signer variability studies.

CLERC builds the data layer underneath sign language AI — not a translation tool, not an accessibility app. Infrastructure.

---

Dataset Summary

• 300 ASL clips — 100 unique phrases × 3 Deaf signers (parallel structure)
• Inter-signer parallel structure — identical phrases across signers for direct variability analysis
• Multimodal keypoints — hands, body, eyes, mouth, head silhouette (MediaPipe-extracted)
• Linguistically validated — ASL gloss annotations with temporal segmentation

This release ships extracted keypoints and annotations only — no raw video. Source clips remain proprietary; access is reserved for commercial licensing (contact florian@clerc.io).

This is v0.1, a pilot release representing a portion of the full CLERC catalog. Full corpus access available via commercial license.

---

Dataset Statistics

Metric	Value
Total clips	300 (100 per signer)
Total frames	36,590
Mean clip length	122 frames (≈ 4.1 s @ 30 fps)
Total signed duration	9.47 min
Gloss tokens	962
Unique glosses	151
Hapax (count = 1)	33 (21.9% of vocabulary)
Mean segments per clip	3.21
MediaPipe head-silhouette detection	100.00% of frames
Frame rate	29.95 – 30.0 fps
Coordinate space	MediaPipe image-normalized (signer perspective)

Top 10 glosses (cumulative coverage of corpus):

#	Gloss	Tokens	% of corpus
1	YOU	180	18.7%
2	QUESTION	151	15.7%
3	WHAT	36	3.7%
4	WHERE	35	3.6%
5	WANT	27	2.8%
6	LIKE	24	2.5%
7	YOUR	22	2.3%
8	HAVE	18	1.9%
9	HOW	14	1.5%
10	GOOD	13	1.4%

---

Languages

• American Sign Language (ASL) — ISO 639-3: ase
• Written translations in English

---

Dataset Structure

epee-v01/
├── keypoints/           # 300 .npy arrays — shape: (n_frames, 128, 3)
├── annotations/         # 300 .json files
└── metadata.csv         # master index (1 row per clip)

Keypoint layout (128 landmarks per frame)

Indices	Region	Source	Notes
0–20	Left hand (21 points)	MediaPipe Hands
21–41	Right hand (21 points)	MediaPipe Hands
42–53	Upper body (12 points)	MediaPipe Pose [11:23]	shoulders, elbows, wrists, finger anchors
54–63	Lower body (10 points)	MediaPipe Pose [23:33]	hips, knees, ankles, heels, feet — spatial context, optional (see below)
64–91	Eyes + mouth only (28 points)	MediaPipe Face	privacy-preserving subset
92–127	Head silhouette (36 points)	MediaPipe FaceMesh FACE_OVAL	forehead, jaw, ears — outline only, no internal features

The 36 head-silhouette landmarks come from MediaPipe FaceMesh FACE_OVAL indices 10, 338, 297, 332, 284, 251, 389, 356, 454, 323, 361, 288, 397, 365, 379, 378, 400, 377, 152, 148, 176, 149, 150, 136, 172, 58, 132, 93, 234, 127, 162, 21, 54, 103, 67, 109 (in that traversal order). The points form a closed polygon outlining the head — no internal facial features are included, so the privacy stance is preserved. Useful for skeleton visualization and as a head-position reference for spatial models.

Coordinate space

Coordinates are MediaPipe's image-normalized space, NOT clipped to [0, 1]:

• x is in [0, 1] (frame width)
• y is in [0, 1] for points visible in frame, but can exceed 1.0 for body landmarks extrapolated below the visible frame (hips, knees, ankles, feet)
• z is depth relative to the hips, roughly in MediaPipe Pose's world-scale units

Source clips are framed waist-up. Lower-body landmarks (dataset indices 54–63) come from MediaPipe Pose's full-body prediction and provide spatial-context anchors for downstream models that benefit from them. For hand/face-only SLR pipelines, they can be dropped:

kp_slr = np.concatenate([kp[:, :54], kp[:, 64:]], axis=1)  # → (n_frames, 118, 3)
# Keeps hands + upper body + face + head silhouette

Zero values (0, 0, 0) indicate a landmark was not detected for that frame (e.g. an off-screen hand).

Gloss conventions

Glosses (uppercase ASL labels) follow a few conventions worth knowing before training:

Base gloss — WHAT, YOU, BATHROOM. The standard form of a sign.

Variants — BASE_N (e.g. SIGN_2, WHAT_3, STUDENT_2). These mark alternative ways to sign the same English concept — different handshape, location, or movement that still maps to the same word/phrase. Across the corpus, about 5% of published clips contain at least one variant gloss. The number N is an internal disambiguator: WHAT_2 is not "more emphatic WHAT", it is a distinct signing form of the same concept. When the base form (e.g. WHAT) also appears in the corpus, treat WHAT, WHAT_2, WHAT_3 as siblings that share the same English target.

Directional / movement suffixes — POINTER_RIGHT, POINTER_LEFT, GO_LEFT, HOW_LEFT_MOVE, HOW_RIGHT_MOVE. These mark spatial/movement components inherent to the sign (_LEFT/_RIGHT/_MOVE). They are not variants and should not be collapsed with their base form.

Phrase repetitions — Some clips contain the target phrase signed more than once (emphasis, demonstration, self-correction). Each occurrence is annotated as a separate gloss segment with its own timestamps. This is natural signer behavior and reflects real inter-signer variability — it is not a labeling error. Users who want strict single-instance training samples can split on segment boundaries.

Recommended preprocessing

# To group variants under one English concept for classification:
import re
def base_gloss(g):
    return re.sub(r"_\d+$", "", g)   # SIGN_2 → SIGN

# To filter clips with phrase repetitions:
from collections import Counter
def has_repeat(segments):
    return any(c >= 2 for c in Counter(s["gloss"] for s in segments).values())

Annotation schema (per clip)

{
  "clip_id": "clerc_v01_101",
  "signer_id": "BRAVO",
  "sign_language": "ASL",
  "text_en": "What's up?",
  "fps": 30.0,
  "n_frames": 139,
  "segments": [
    { "gloss": "WHAT'S UP", "start": 0.9, "end": 1.4 },
    { "gloss": "WHAT",      "start": 1.6, "end": 1.8 },
    { "gloss": "QUESTION",  "start": 2.0, "end": 2.8 }
  ]
}

---

Signers

signer_id	Gender	Age range	Language acquisition	Clips
ALPHA	F	30–40	Native Deaf signer (ASL L1)	clerc_v01_001 → 100 (100 clips)
BRAVO	M	30–40	Native Deaf signer (ASL L1)	clerc_v01_101, 103, 105 … 299 (100 clips)
CHARLIE	M	30–40	Native Deaf signer (ASL L1)	clerc_v01_102, 104, 106 … 300 (100 clips)

Demographic distribution: 1 female / 2 male, all between 30–40 years old, all native ASL signers (Deaf, ASL as first language). Signer identities are pseudonymized.

Signers participated under written informed consent. The signing space, framing, lighting, and recording protocol were standardized across signers.

Parallel structure: all three signers sign the same 100 phrases, enabling direct inter-signer comparison. ALPHA's clips (001–100) follow phrase order; BRAVO and CHARLIE alternate in clips 101–300 (BRAVO on odd indices, CHARLIE on even).

Stylistic note: Phrase repetition appears in roughly 22% of ALPHA clips and ≤3% of others — natural inter-signer stylistic variation, annotated as separate gloss segments. See gloss conventions for filtering.

---

Intended Use

Designed for

• Inter-signer variability analysis (style, rhythm, signing space)
• Research on sign language linguistics, gesture recognition, multimodal AI
• Educational use in academic settings
• Prototyping sign language recognition (SLR) pipelines on a parallel multi-signer corpus

Not designed for

• Speaker identification or biometric applications
• Surveillance or evaluation of individual signers

For production-grade systems or sign language generation models trained at scale, see commercial licensing for access to the full multi-signer corpus.

---

Loading the Dataset

This release ships as plain .npy + .json files for transparency and zero-dependency loading.

Note: the datasets library's load_dataset() is not the right entry point here — this dataset uses raw NumPy arrays (not Parquet/Arrow) and there is no loading script, by design. Use huggingface_hub.snapshot_download() to fetch all files locally, then load with numpy + pandas as shown below.

import json
import numpy as np
import pandas as pd
from pathlib import Path
from huggingface_hub import snapshot_download

ROOT = Path(snapshot_download(repo_id="CLERC-DATA/epee-v01", repo_type="dataset"))

# Load metadata
metadata = pd.read_csv(ROOT / "metadata.csv")
print(metadata.head())

# Load one clip's annotation + keypoints
clip_id = "clerc_v01_101"
with open(ROOT / "annotations" / f"{clip_id}.json") as f:
    annotation = json.load(f)

keypoints = np.load(ROOT / "keypoints" / f"{clip_id}.npy")
print(annotation["text_en"], keypoints.shape)
# → What's up? (139, 128, 3)

# Convenient slices:
hands       = keypoints[:, :42]      # both hands (42 pts)
upper_body  = keypoints[:, 42:54]    # shoulders → wrist anchors
face_inner  = keypoints[:, 64:92]    # eyes + mouth
head_oval   = keypoints[:, 92:128]   # head silhouette

---

License

CC BY-NC-SA 4.0 — creativecommons.org/licenses/by-nc-sa/4.0

You are free to:

• Share — copy and redistribute
• Adapt — remix, transform, build upon

Under these terms:

• Attribution — give credit, link to license, indicate changes
• NonCommercial — no commercial use
• ShareAlike — distribute contributions under same license

Commercial licensing: for enterprise use, training of commercial models, or integration into commercial products, contact florian@clerc.io.

---

Ethical Considerations

CLERC is Deaf-led infrastructure. This release adheres to the following principles:

• Informed consent — all signers have provided written consent for public release of their data under this license
• Privacy protection — face landmarks are restricted to non-identifying features (eyes + mouth); full biometric data is excluded
• Community benefit — the dataset is released to advance sign language technology research; commercial revenue supports continued Deaf-led data infrastructure
• No surveillance use — this data must not be used for individual identification, behavioral profiling, or any application that surveils or evaluates individual signers

If you have concerns about the use of this dataset, contact florian@clerc.io.

---

Limitations

• Pilot release — 300 clips is a baseline pilot, not a production-scale corpus
• 3 signers — limited inter-signer diversity; full catalog includes broader signer pool
• Phrase domain — focused on conversational/social phrases; not domain-specific (medical, legal, technical)
• Reduced face landmarks — full facial grammar (brow, cheeks, head tilt) not included in this release
• Gloss only — no morphological, prosodic, or spatial annotation layers in v0.1

These limitations are intentional for the v0.1 release. Full multi-layer annotations available via commercial license.

---

Versioning & Roadmap

Version	Status	Content
v0.1	✅ Current	300 clips, 3 signers (ALPHA, BRAVO, CHARLIE), gloss + timing
v0.2	Planned Q3 2026	Expanded signer pool, additional categories
v1.0	Planned 2027	Multi-layer annotations, broader corpus

---

How to Cite

@dataset{clerc_epee_v01_2026,
  author       = {M{\'e}loux, Florian and {CLERC}},
  title        = {{CLERC} {\'E}p{\'e}e v0.1: Sign Language Data Layer},
  year         = {2026},
  publisher    = {Zenodo},
  version      = {0.1},
  doi          = {10.5281/zenodo.20268568},
  url          = {https://doi.org/10.5281/zenodo.20268568},
  note         = {CC BY-NC-SA 4.0. Mirrored at https://huggingface.co/datasets/CLERC-DATA/epee-v01}
}

---

About CLERC

CLERC builds the data infrastructure that lets AI understand sign language as a first-class language — not an accessibility afterthought.

Sign language is not to be translated. It is to be inscribed.

Website: clerc.io Manifesto: clerc.io/manifesto Contact: florian@clerc.io LinkedIn: clerc-io

---

Changelog

v0.1 — May 2026

• Initial public release
• 300 ASL clips, 3 signers (ALPHA, BRAVO, CHARLIE), parallel structure
• 128 multimodal keypoints per frame (hands + body + eyes/mouth + head silhouette)
• Gloss annotations with temporal segmentation