# Épée v0.4 · Signer-Robustness Benchmark

**What it shows:** signer variation is the core problem in sign-language AI, and the fix is
*more signers and more data*, not a bigger model. Re-measured on the eight v0.4 signers with the
v0.2 / v0.3 protocol, **with one deliberate change: hand slots are now consistent in train and
test.**

The model is trained on release signers and tested on **two native Deaf signers with no clips in
this release** (Signer A and Signer B, pseudonymized, from the broader CLERC corpus). They recorded
under the same protocol and share prompts with the grid signers: same content, brand-new body,
the controlled comparison a signer-generalization test needs. No clip is shared between train and
test.

![benchmark](benchmark.png)

---

## Why the v0.3 numbers are not comparable

v0.1 to v0.3 shipped with crossed hand slots: the `0-20` / `21-41` blocks were not consistently
the same hand, and the convention differed by signer (measured on the published v0.3 files: most
decidable clips crossed). A model trained on that learns "which slot holds the active hand" as a
signer cue, and a stranger whose slots go the other way is penalized for it. v0.4 fixes the
release at build time and this run fixes the frozen test segments with the same rule, per sample:
Signer A needed 10 of 1,961
segments swapped, Signer B 1210 of
1,453. Fixing only the test side would be an exploit, not a correction;
both sides are fixed here.

So the effects are separated, all at 8 seeds:

| Training data | Signer A | Signer B |
|---|---|---|
| v0.3 as published (crossed slots, train and test) | 69% ± 5 | 58% ± 6 |
| v0.3 recordings, slots fixed (same 1,199 clips) | 76% ± 4 | 60% ± 4 |
| v0.3's six signers at 300 clips each (2,477 segments) | 77% ± 3 | 63% ± 9 |
| all eight signers, subsampled to the same 2,477 segments | 79% ± 4 | 64% ± 6 |
| **v0.4, eight signers, 2,400 clips (2,877 segments)** | **81% ± 1** | **69% ± 4** |

The first step is the correction alone (same clips). The second is more clips for the same six
signers. The third row holds the volume fixed and only changes the signer mix: at equal volume the difference is +2.3 points on Signer A and +1.4 on Signer B, within run-to-run spread (±4): this run does not separate diversity from volume.
The last row adds the remaining volume.

---

## Setup

| | |
|---|---|
| Model | small BiLSTM (1 layer, hidden 64, bidirectional, mean-pool → linear) |
| Input | keypoint sequences, hands + arms (48 pts, x/y), shoulder-normalized, resampled to 24 frames |
| Task | sign classification on the **same 24 glosses** as v0.2 and v0.3 (vocabulary frozen) |
| Train | 2,877 gloss segments from the eight release signers (1,393 in v0.3) |
| Test | Signer A: 1,961 segments · Signer B: 1,453 segments (same as v0.2 / v0.3) |
| Chance | 4.2% (1 / 24) · **8 seeds** per configuration, mean ± std |
| Metrics | top-1 accuracy **and macro-F1** |

## Result 1 · More signers → better on a brand-new signer

Tested on Signer A. Mean over signer combinations at each size: every single signer at k = 1, then
a seeded sample of combinations (C(8, k) reaches 70; the sample is an unbiased estimate of the
all-combinations mean reported in v0.2 and v0.3).

| # training signers | combinations | Accuracy | Macro-F1 |
|---|---|---|---|
| 1 | 8 | 42% ± 15 | 0.28 |
| 2 | 5 | 64% ± 12 | 0.52 |
| 3 | 5 | 67% ± 6 | 0.57 |
| 4 | 5 | 73% ± 8 | 0.65 |
| 5 | 5 | 75% ± 7 | 0.67 |
| 6 | 5 | 77% ± 4 | 0.69 |
| 7 | 5 | 79% ± 5 | 0.71 |
| 8 | 1 | 81% ± 1 | 0.74 |

A signer the model *has* seen scores 83% (ceiling). At eight training signers the stranger
gap is **1 point**. Second stranger, Signer B: 69%.

The ± is the spread over seeds and combinations. At k = 8 there is a single combination, so its ±
is seed variance only: the narrowing toward k = 8 is partly a counting effect, not only a
stabilization of the model.

## Result 2 · More data → still rising, more slowly

Same held-out signer; training set grown by a clean random fraction of the v0.4 pool:

| Training examples | 80 | 160 | 260 | 380 | 520 | 680 | 858 | 1,044 | 1,393 | 2,157 | 2,877 |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Accuracy | 42% | 61% | 56% | 70% | 70% | 73% | 72% | 75% | 76% | 77% | 80% |

---

## Caveats

- 24-gloss shared subset, frozen from v0.2 for comparability. It is a measurement instrument,
  not a ceiling on what the release covers (1,501 unique glosses in v0.4).
- **Dominant hand is not normalized.** Hand slots are fixed to the left / right of the body, they do
  not swap dominant and non-dominant hands. Alone, signers range from 14% (ALPHA) to
  62% (CHARLIE) on Signer A. The low end is ALPHA, whose dominant hand is on the opposite side
  of the frame from Signer A's (it was already the low end in v0.3), not the own-register signers.
  Signer B is in the opposite case to Signer A, which is consistent with its lower score. Mirroring
  left-dominant signers is a modelling choice left to the user.
- FOXTROT, GOLF, HOTEL cover fewer than the 24 test glosses (FOXTROT 18, GOLF 15, HOTEL 18):
  their prompt sets are other registers (declarative sentences).
- The full-pool point of the data curve (80%) and the headline v0.4 number (81%)
  train on the same segments in a different order (the curve shuffles before slicing); the gap is
  run-to-run noise, not a different setting.
- Measures generalization across signer **bodies**, not across topics.
- Small BiLSTM by design; the 1-signer vs many-signer gap is a data property.
- 26 Signer A
  and 8 Signer B
  test segments had too little hand evidence to decide their slots and are left as extracted.