Ope0/vision-gap

Yoruba Vision Gap

Small VLMs read the letters of Yoruba but not its meaning — and bend Nigerian documents toward Western defaults

Model evaluated: Qwen/Qwen3-VL-4B-Instruct (≈4.4B parameters, Apache-2.0) · optional second model Qwen/Qwen2.5-VL-3B-Instruct Evaluation set: 360 synthetic, exactly-labelled probes (352 images) in 6 families, every probe paired with a control Everything here is reproducible: vision_gap_eval.ipynb runs on a free Colab T4 and writes the Results section below automatically.

Path	What it is
README.md	this file — answers to Q1 (the gap) and Q3 (path forward), method, results
data/images/	352 PNG/JPG probe images
data/metadata.jsonl / .csv	one row per probe: id, family, condition, prompt, ground truth, answer type, options, metadata
build_dataset.py	deterministic generator for the whole set (python build_dataset.py --out data --seed 7)
scoring.py	diacritic-aware automatic scoring (exact, letters-only, mark recall/precision by mark type, ANLS for comparison)
vision_gap_eval.ipynb	Colab notebook: load model → run → score → tables → figures → README → upload
results/<model>/predictions.jsonl	every raw model answer, with latency
results/scored.csv, results/summary.md, results/failure_gallery.md, results/figures/	scored rows, headline table, worst cases, plots
model/<model>.json	pinned model revision (commit SHA), dtype and decoding settings used

---

1. The blind spot (Question 1)

What it is

Yoruba is a tonal language written in the Latin alphabet. Its meaning lives in two kinds of marks that standard English OCR treats as decoration:

• tone marks — grave (low tone, à), acute (high tone, á), unmarked (mid tone);
• under-dots — ẹ, ọ, ṣ are different letters from e, o, s.

Take away the marks and words collapse into each other:

written	meaning	written	meaning
ọkọ	husband	igbá	calabash
ọkọ̀	vehicle, boat	ìgbà	time, period
ọ̀kọ̀	spear	igba	two hundred
ọkọ́	hoe	ìgbá	garden egg

The gap I am testing is that vision-language models read the base letters of Yoruba text correctly but drop, flatten or invent the marks, so they return a different word while looking almost right. Around it sits a broader pattern I see in Nigerian documents generally: when local visual evidence conflicts with the model's Western prior, the prior wins. The naira sign ₦ becomes N or $; 03/04/2026 on an invoice becomes March instead of April; and a model asked "the total on this Bodija receipt is $13,200, correct?" is tempted to say yes.

Why this comes from my own background

• Signage and names. Street signs, church and mosque boards, market stalls, palace and council notices, and people's names (Adébáyọ̀, Fọláṣadé, Olúwaṣẹ́un) are written with tone marks. Strip them and an identity or a place name is wrong — the kind of error that breaks a student-records system, a KYC check or a land document search.
• Documents arrive as WhatsApp photos. Receipts, admission letters, bank tellers and government circulars are photographed on a phone and forwarded several times. Each hop recompresses the JPEG, and tone marks — two or three pixels tall — are the first detail to disappear.
• Money and dates are local. Prices are in naira; on a keyboard without ₦ people write N2,500 or #2,500. Dates are DD/MM/YYYY. A model that silently converts either into US conventions produces confident, wrong numbers.

Why standard benchmarks miss it

1. Coverage. Text-in-image benchmarks (TextVQA, DocVQA, ChartQA, OCRBench, and multilingual ones such as MTVQA and CC-OCR) are dominated by English and a handful of high-resource languages; to my knowledge none tests Yoruba orthography, the naira sign or Nigerian document conventions. Culturally diverse VQA sets ask about scenes, not about reading tone-marked text.
2. The metric hides it. OCR benchmarks score with character error rate or ANLS. A dropped tone mark is a single character edit: reading Ẹ kú àárọ̀ ("good morning") as Ẹ kú àárọ gets ANLS 0.90 — a pass and reading ọkọ̀ (vehicle) as ọkọ (husband) gets 0.75. Some evaluation pipelines also strip accents before comparison, which erases the error entirely. The gap is therefore invisible by construction, not just by omission.
3. Leaderboards reward the prior. On English-heavy test sets, defaulting to $ and MM/DD is usually right, so the behaviour that fails Nigerian users is the behaviour that scores well.

---

2. Choice of model and how it was evaluated (Question 2)

Model: Qwen/Qwen3-VL-4B-Instruct

• Size: ≈4.4B parameters — inside the 0.6–6B window and runs in fp16 on a free 16 GB T4.
• Frontier-representative: it shares the architecture and training recipe of the Qwen3-VL family whose 235B flagship is among the strongest open-weight VLMs, so its failure modes are a fair, cheap proxy for the flagship's (the notebook includes an optional cell that re-asks the hardest slice to the 235B model to check this directly).
• A fair test, not a strawman: the model card advertises expanded OCR for 32 languages, robustness to "low light, blur, and tilt", native dynamic resolution, and DeepStack multi-level vision features for fine detail. If a model that claims all of this still drops tone marks, the problem is not image resolution or a weak OCR head.
• Second opinion: Qwen2.5-VL-3B-Instruct (previous generation) can be switched on in the notebook to test whether the gap persists across generations; Gemma-3-4B and SmolVLM2-2.2B are also supported by the same harness.

Decoding is greedy (do_sample=False), one image per prompt, no system prompt, so results are deterministic and reproducible. The exact model revision is pinned in model/*.json.

The evaluation set

All 360 probes are rendered synthetically (Pillow + HarfBuzz/raqm, so stacked marks like ọ̀ are positioned correctly) from five fonts and five sign colour schemes. Synthetic rendering gives exact ground truth and lets every probe change one variable at a time.

Family	n	Probe	Matched control	Question it answers
diacritics / yo	96	24 Yoruba items (12 minimal-pair words, 6 place names, 6 phrases) × 4 conditions	—	Can it transcribe Yoruba exactly?
diacritics / yo-ascii	48	the same 24 items with all marks removed (2 conditions)	same letters, same font, same layout	Is the failure about the marks, or about the words? Does it invent marks?
diacritics / vi	32	8 Vietnamese items × 4 conditions	same mark shapes (dot-below, grave, acute, stacked) in a high-resource language	Is it the pixels, or the language?
diacritics / fr	32	8 French items × 4 conditions	grave/acute in the highest-resource accented language	—
minimal_pairs / mc_spelling	24	pick the exact spelling from the minimal set (+ stripped form)	chance ≈ 23%	Can it see the mark when it doesn't have to write it?
minimal_pairs / meaning_from_image	8	"what does this word mean?" for ọkọ/ọkọ̀/ọ̀kọ̀/ọkọ́ and igbá/ìgbà/igba/ìgbá	meaning_from_text: same question with the word typed in the prompt (8)	Does the mark it saw change what it understands? Does it know the word at all?
currency / symbol_*	32	price tags ₦500 … ₦15,000, clean and WhatsApp-compressed	identical tags in $, £, €	Is ₦ read as ₦?
currency / name_*	12	name the currency: ₦ glyph; N2,500 and #2,500 on a Bodija Market tag	—	Does it know Nigerian price conventions?
receipts	24	Bodija provision-store receipt in ₦: neutral total + false-premise total (wrong currency or wrong amount)	identical Austin, TX receipt in $	Is it more sycophantic on local documents?
dates	16	invoice dated 03/04/2026-style (day ≤ 12)	the same strings on US invoices; invoices with day > 12	Does it read the date in the document's own locale?
known_controls	28	counting 3–32 dots, left/right, English menu transcription + false premise	—	How does this gap compare with blind spots already in public benchmarks?

Conditions (diacritics family): L_clean 72 px text, lossless · L_wa same, then WhatsApp-forwarded (±2.5° tilt, 0.6× downscale, slight blur, JPEG q=35) · S_clean 28 px text · S_wa 28 px + WhatsApp.

Scoring (scoring.py)

• Exact match after Unicode NFC, case-folding and whitespace cleanup — tone marks and dots are never stripped.
• Letters-only match (marks removed on both sides) — separates "read the wrong letters" from "read the right letters, wrong marks" (marks_only_error).
• Mark recall / precision by type (dot_below, grave, acute, …) after aligning base letters, plus spurious marks on the stripped control (hallucinated diacritics).
• ANLS on the same answers, to show what a standard OCR benchmark would have scored.
• Currency symbol and amount parsed separately (N2,500 is reported as semantically right but not exact); YES/NO parsing for false premises, plus whether the true total was stated; month parsing for dates; letter parsing for multiple choice.

How the probes add up to a diagnosis

If the results show…	…the gap is
Yoruba ≪ stripped Yoruba, and Yoruba ≪ Vietnamese on the same mark types	about the language, not about seeing small marks (data coverage)
multiple-choice spelling ≫ free transcription	a generation prior: the model sees the mark but writes the ASCII form
multiple-choice ≈ chance and falls with font size / compression	a perception limit on 2–3 px features
meaning-from-text ≫ meaning-from-image	the model knows the word but the visual mark doesn't reach its reasoning
₦ ≪ $ £ €, US-reading of Nigerian dates, higher sycophancy on the Bodija receipt	a locale prior overriding visual evidence
known controls fail too, but Yoruba fails at sizes where ASCII succeeds	the gap is specific, not general weakness

---

3. Results

Pending: run vision_gap_eval.ipynb (cell 10 writes this section from the run).

---

4. Path forward (Question 3)

The probes above are designed to say which of the fixes below matters most; this is the plan I would follow.

4.1 Data curation — the main lever

1. Synthetic Yoruba text-in-image at scale (cheap, exact labels). Scale build_dataset.py into a renderer driven by real Yoruba text — MENYO-20k, Yoruba Wikipedia, BBC Yorùbá news, public-domain religious text — rendered in 50+ fonts including hand-painted sign styles, onto signs, receipts, posters, bulletins and government forms, with a degradation pipeline that reproduces how Nigerian documents actually travel: repeated WhatsApp JPEG compression, phone-camera blur, perspective, low light. Rendering must go through HarfBuzz (as here) or stacked marks come out wrong and poison the data. Target: ~1M image–text pairs, NFC-normalised.
2. Real photos with native-speaker labels (small, high value). Photograph signage, receipts and notices in Nigeria and have two native speakers transcribe each one, with disagreements adjudicated. A few thousand images is enough for evaluation and for a final fine-tuning stage; it also catches things synthesis misses, such as hand lettering and faded paint.
3. Minimal-pair hard negatives. For every image, store alternative transcriptions that differ only in marks (ọkọ / ọkọ̀ / ọ̀kọ̀ / ọkọ́). These become the rejected side of preference pairs and the distractors of contrastive training — the model must learn that these differences matter.
4. Locale-grounded document QA. Nigerian receipts, invoices, tellers and forms with ₦, N and # prices, DD/MM dates and +234 numbers, paired with question–answer data in which a good share of questions carry a false premise whose correct answer is "No, it says …".
5. Fix the measurement. Report a diacritic error rate and mark recall separately from CER/ANLS, never strip accents during evaluation, and add Yoruba and Igbo (ị, ọ, ụ, ṅ) slices to OCR leaderboards. As long as the metric scores ọkọ̀ → ọkọ at 0.75, no one will optimise for it.

4.2 Fine-tuning paradigm

1. Stage 1 — supervised fine-tuning that reaches the vision side. LoRA on the language model plus unfreezing the vision–language merger and the top vision blocks (for Qwen3-VL, including the DeepStack fusion paths), trained on transcription and document QA from 4.1. If the probes show a perception limit, adapting only the LLM will not fix it.
2. Mark-weighted loss. Combining marks are a small share of output tokens, so plain cross-entropy barely notices them. Up-weight the loss on tokens containing tone marks or under-dots (e.g. 3–5×).
3. Stage 2 — preference optimisation. DPO (or an image-conditioned variant such as mDPO) with pairs chosen = correct marks vs rejected = stripped or wrong tone, and chosen = "No, the total is ₦13,200" vs rejected = "Yes". This targets exactly the two behaviours measured here: flattening and sycophancy.
4. Keep everything else intact. Mix in ≥50% general replay data and re-run OCRBench and the known controls after each stage so that fixing Yoruba does not break English.

4.3 Architectural adjustments

1. Resolution where the text is. A tone mark on 28 px text is 2–3 pixels; after patching and 2×2 token merging one visual token covers roughly a 32×32 px area, so the mark is a small fraction of one token. Detect text regions and re-encode them at higher resolution (a learned "zoom" pass), rather than paying for high resolution everywhere.
2. Character-level grounding head. Add an auxiliary CTC OCR head on the vision features, trained on diacritic-exact targets, so the model has pixel-grounded evidence for each mark that the decoder can attend to.
3. Tokenizer. Byte-level BPE splits ọ̀ into several rare tokens, which gives the decoder a strong prior toward the ASCII spelling. Adding tokens for frequent Yoruba vowel-plus-mark clusters (initialised from the mean of their pieces) lowers that prior.
4. Calibrated abstention. When the evidence for a mark is weak — a heavily compressed photo — the model should say the tone is unreadable instead of guessing; this is cheap to train with the hard negatives above.

Cheapest test of this plan: LoRA-tune Qwen3-VL-4B on ~50k synthetic renders from 4.1 (stage 1 + mark-weighted loss), then re-run this notebook. Success = Yoruba exact match reaches the Vietnamese control on held-out fonts and real photos, with no drop on OCRBench and the known controls.

---

Limitations

• Images are synthetic. They control variables cleanly but do not capture hand lettering, faded paint or real camera optics; the real-photo set in 4.1 is the natural next step.
• The ambiguous-date probes have no single "correct" answer in isolation; they measure whether the model uses the locale cues present (Addresses, +234 phone, ₦ amounts), and are reported separately from unambiguous dates.
• Yoruba orthography follows standard written Yoruba and the minimal-pair glosses were reviewed by the author; please open an issue if you spot an error.
• Greedy decoding on one prompt wording per probe; prompt sensitivity is not measured here.

Licence and citation

Dataset (images + metadata): CC-BY-4.0. Code: MIT. Model weights: see the original model licence (Apache-2.0).