k-imtz/vl-blindspot-indian-admin-docs

VLM Blind Spot: Indian Administrative Scans

FIT 2026 fellowship technical challenge.

Model: Qwen/Qwen2.5-VL-3B-Instruct, 4-bit (nf4, fp16 compute), greedy decoding. Data: real property-affidavit scans from bpsm.bihar.gov.in (Bihar govt, public records). Evaluator: Md. Kaif Imteyaz, Jamia Hamdard / Michigan State University.

---

Q1: Blind spot

I work on CivicBot, a pipeline that extracts asset data from 65,344 scanned Bihar government officer affidavits. Each affidavit is 3 pages. Page 1 has identity fields (name, designation). Pages 2 and 3 have asset tables (property, cash, loans, PAN).

These scans are mixed Hindi/English, with handwritten fields over printed forms. Standard VLM benchmarks do not cover this kind of document.

The model does not invent values it cannot read. It returns null. The problem is that null looks identical to a blank-on-form field, so the pipeline sees a valid JSON response with no signal that extraction actually failed.

The failure is page-structure driven. Page 1 (identity) extracts at 73%. Pages 2 and 3 (asset tables) extract at 0%. The model handles simple labeled fields but cannot parse dense tabular asset data.

---

Q2: Evaluation

33 pages across 33 PDFs from the Bihar BPSM portal. Fields: officer_name, designation, district, tehsil, khasra_no, prop_value. No pre-processing. Every page sent identically: image + fixed prompt.

Failure taxonomy:

• extracted - all 6 fields present, at least one non-null
• partial_refusal - 3 or more fields null
• all_null_refusal - every field null
• textual_refusal - prose refusal instead of JSON
• schema_missing_fields - JSON parsed but keys missing
• parse_failed - output not parseable as JSON

Results:

Failure mode	Count	%
partial_refusal	18	54.5%
extracted	8	24.2%
all_null_refusal	7	21.2%
n = 33 pages across 33 PDFs.

By page:

Page	Contents	extracted %	partial_refusal %	all_null %
1	Name, designation, identity fields	72.7	27.3	0.0
2	Asset declarations (movable/immovable)	0.0	36.4	63.6
3	Asset continuation + PAN/income tax	0.0	100.0	0.0

No ground truth labels. Failure flags verified by human spot-check on a sample.

---

Q3: Path forward

1. Page-type routing. Identity fields (page 1) and asset tables (pages 2-3) fail at very different rates. Route by page structure and apply different confidence thresholds or fallback strategies per page type.

2. Test-time augmentation. Run inference on the raw page plus a deskewed and contrast-enhanced version. Take the output with the most non-null fields. No fine-tuning needed.

3. Refusal supervision. Fine-tune on page pairs where null is the correct answer for genuinely unreadable fields. The model needs to learn that null means field-level unreadability, not document-level confusion.

---

Privacy

Source PDFs are public records but contain officer names and property details. This repo publishes only blurred thumbnails, failure flags, and aggregate counts. No field values. Raw PDFs never leave the eval machine.

Artifacts

File	What it is
eval_results_redacted.json	Per-page flag + JSON shape. All values [REDACTED].
failure_mode_counts.csv	Aggregate flag distribution.
eval_images_redacted/	Blurred thumbnails with black bars over text regions.

Reproduce

Notebook: FIT_Eval_Qwen3VL_4B.ipynb. Requires: transformers>=4.49,<4.57, accelerate, bitsandbytes, pillow, pdf2image, huggingface_hub, pandas, poppler. Upload target_pdfs_latest.csv from the CivicBot pipeline when prompted.

CivicBot: https://civicbot-seven.vercel.app