FROM HANDWRITING TO EVIDENCE

A model, measured.

Follow the Telugu word recognition pilot from verified data to predictions. Every result here comes from the local run on your computer.

Connecting to local study…

Training words1,000One epoch · LoRA rank 8

Validation words500Separate source-document groups

Held-out test words500Same inputs for both models

LIVE RUN

Loading study status

Live status is available from the local API. The hosted page can show the last exported snapshot.

SAME TEST SET · GREEDY DECODING

Original weights vs. Telugu adapter

Measure	Original IndicOCR	Word adapter
Waiting for completed test evaluation. No accuracy claim yet.

NFC grapheme character error rate (CER) counts substitutions, insertions and deletions. Lower CER and WER are better; higher exact match is better. Punctuation and digits stay in the comparison.

REAL INPUTS · UNEDITED OUTPUTS

Read alongside the model

Test examples will appear as the adapted model is evaluated. These are the first examples in the frozen evaluation order, not selected successes.

What this study establishes

Original Qwen3.5 recognizer and shipped Telugu-capable tokenizer.
Image and label verification, duplicate grouping, and frozen split checksums.
Finite gradients, changed adapter weights, tiny-subset learning and save/reload checks.
Word recognition on a bounded pilot, using source-provided ground truth.

What remains to be measured

Whole-page handwriting and layout accuracy.
Writer-disjoint generalization; writer identities are unverified.
Human review of source labels and near-duplicate groups.
Robustness to camera photos, faint ink and complex layouts.

Local research workspace · source data stays on this computer · local results refresh every 15 seconds · hosted results are a saved snapshot

Handwriting examples: IIIT Hyderabad, IIIT-INDIC-HW-WORDS Telugu · source · CC BY 4.0. Original image bytes preserved; displayed at scaled size. Live model inference is local.