Bittensor SN117 · OCR Research

Glyph

A rolling-winner benchmark for lightweight OCR models that read private synthetic documents with speed, accuracy, and restraint.

Launch dashboard Join the subnet

invoice total 48.20 uid 117

plain text output INVOICE #117 TOTAL 48.20 CONFIDENCE 0.94

Glyph Research operates an OCR-only subnet where miners commit one public HuggingFace model and validators score it against unpublished English document challenges.

Benchmark

Private OCR tasks, public winners.

70%accuracy

Normalized character and word edit distance.

15%latency

Throughput under a single-validator GPU budget.

10%efficiency

Model size and peak memory pressure.

5%robustness

Rotation, blur, compression, font variance, and noisy captures.

Protocol

Rewards concentrate around proven winners.

current winner60%

previous winner30%

previous previous10%

Validator state tracks committed revisions, model hashes, prechecks, benchmark scores, and disqualification reasons. Historical winners are compacted when a model becomes invalid.

Participate

Commit a model. Let the benchmark decide.

01 Read the commit format

Use a pinned HuggingFace revision and standard Transformers loading.

02 Check model constraints

Public repo, safetensors, image input support, no remote code.

03 Watch the leaderboard

Only eligible rolling winners receive non-zero weights.