Lab Update: Reading Faded Ink

“All that you touch You Change. All that you Change Changes you. The only lasting truth is Change.”

Octavia E. Butler

Wuh gine on! Anaiya here. 👋🏾

It has been an intense week in the lab. ⏱️ We jumped headfirst into the R.O.A.D. Barbados Historic Handwriting Challenge on Zindi, an initiative aimed at transcribing thousands of degraded 18th- and 19th-century Barbadian archival records.

When you look at these colonial-era survey ledgers and legal deeds, you quickly realize how tough the challenge is. You are dealing with historic secretary hand: multiple scribes with wildly different cursive flourishes, faded iron gall ink, water stains, bleed-through from the back of the page, and severe line overlap. Off-the-shelf optical character recognition baselines struggle significantly on documents like these, often producing character error rates exceeding 60%.

When we ran our initial zero-shot vision baselines, our scores were stuck in the low 0.30s. The models were hallucinating modern English phrasing instead of deciphering archaic abbreviations and orthography.

We locked in. 🔬

1. Fixing the Metric Trap: We caught an inverted scoring trap early where we were calculating error rate inversely. Zindi’s official leaderboard score is normalized as 1 − (0.5·wCER + 0.5·wWER) where higher is better. Aligning our local 410-line holdout to match the exact competition metric brought our local validation within 0.005 of our subsequent leaderboard submission.
2. Vision-Tower LoRA: We discovered that standard PEFT starter configs only targeted language attention projections (q_proj), completely bypassing Qwen’s visual modules (proj, fc1, fc2). Enabling vision-tower adaptation contributed to a +0.064 jump on the public leaderboard (as recorded in our July 2026 submission logs).
3. Dual-Engine Strategy: We deployed a two-pronged architecture: fine-tuning Qwen2-VL-2B on the language and vision side, while training Kraken CTC (fine-tuned from the McCATMuS historical base) inside a WSL2 Ubuntu environment.

That combination propelled us from rank 121 up to rank 59 with a public leaderboard score of 0.8717 (July 16, 2026 submission snapshot).

This sprint also forced us to automate our workspace discipline: strict holdout isolation, deterministic preprocessing transforms, and reusable challenge scaffolding scripts.

We are building the tools that build the tools. 🛠️

Talk soon,
Anaiya ✨