TELUGU HANDWRITING · COMPLETED LOCAL STUDY
Our OCR,
measured.
One capped training epoch. An untouched final set until selection. A direct comparison with the same original IndicOCR checkpoint.
Loading the verified results snapshot…
IDENTICAL INPUTS · DETERMINISTIC DECODING
Original weights. Pilot. Selected adapter.
The candidate was locked using validation only. Final outcomes did not change that choice. This ranking is descriptive of these three variants on the same 3,000 word crops.
| Variant | Crops | Raw CER | NFC CER | Raw WER | NFC WER | Raw exact | NFC exact |
|---|
Swipe tables horizontally on a narrow screen. Lower CER/WER and higher exact match are better.
TRAIN—One effective epoch · 6,250 updates
VALIDATE—Initialization and candidate selection
TEST—Sealed until selection · same inputs
PAIRED EVIDENCE
Improvement, with uncertainty.
2,000 bootstrap repetitions, seed 42, across 128 source-document / duplicate components. These groups are not verified independent writers. Positive error reductions and exact gains favor the already-selected model.
| Original → selected | Difference (pp) | Paired 95% interval (pp) |
|---|
Input-defined subgroup results
Original versus selected. Output-dependent error-flag groups are excluded from this matched subgroup table.
| Subgroup | N | Original CER | Selected CER | Original WER | Selected WER | Original exact | Selected exact |
|---|
Output diagnostics and resource measurements
| Variant | Median generation | p95 | Peak allocated | Repetition flags | Truncations |
|---|
Latency measures generation, excluding upload/HTTP and input transfer. Original peak VRAM covers its remaining 773-example process; pilot and selected peaks cover their complete 3,000-example passes. Hardware and thermal conditions varied.
PUBLISHED CONTEXT · CHECKED 15 OCTOBER 2026
Benchmarks have boundaries.
External tables below use their publishers’ datasets, units and protocols. Akshara was not evaluated on those exact external test sets. Scores from different tables cannot form one worldwide leaderboard.
IndicOCR-HW · Telugu
Vendor score: 100 × (1 − WER). Model-card source ↗
| Order in this row | Model | Score ↑ |
|---|
Akshara: not evaluated on this benchmark. Its 76.20% local exact match must not be inserted into this table.
IndicOCR’s published benchmark families
| Published task | IndicOCR | Unit |
|---|
Vendor H100 80GB page pipeline: 6.34 pages/s at concurrency 128. Our RTX 3050 word timings measure a different workload. Model-card results ↗
OmniDocBench v1.6_full · published top five
Page-parsing composite across text, formulas and tables; higher is better. This is not Telugu word exact match. Table version retained as labeled by the repository. Official benchmark repository ↗
| Order by this score | Model | Overall ↑ |
|---|
ICDAR 2024 HWD · Telugu isolated words
Competition teams, ranked here by published Telugu WRR. Different test inputs and recognition-rate definitions; these are not our grapheme-CER results. “IndicOCR Group” is a competition team name, not an identification of bodhan-ai’s checkpoint. Organizer leaderboard ↗
| Order by WRR | Team | CRR ↑ | WRR ↑ |
|---|
Published Telugu HTR · in-domain and unconstrained
Paper-reported WER/CER on its IIIT dataset splits. Training budgets, unlabeled-data use and evaluation definitions differ from this capped study. No shared Akshara rank follows. SemiHastakshar paper, Tables 4–5 ↗
| Domain | System | WER ↓ | CER ↓ |
|---|
REPRODUCIBILITY & LIMITS
Keep the evidence with the model.
What is verified
- Original 31-file checkpoint and frozen manifest hashes preserved.
- Native BF16 recovery smoke reproduced uninterrupted adapter weights.
- Separate latest resumable, best-validation and versioned selected artifacts.
- Checksum/ZIP integrity, 85 software tests, actual 64/64 API parity and headed desktop/mobile checks.
What remains unknown
- Writer identity and original-model pretraining overlap.
- Complete human semantic verification of source labels.
- Other datasets, camera-image robustness and whole-page accuracy.
- Between-seed robustness; this study used one seed and one capped epoch.
Selected output still has 714 non-exact final crops. Diagnostic flags are heuristics, not calibrated confidence. Constant-pixel abstention does not prove general blank-image detection. Backups are verified locally on D:; the original study did not establish an offsite backup. Hosting this report does not deploy GPU inference.