Datasets:

mvbalaji
/

tamil-ocr-benchmark

text stringlengths 0 89
# ============================================================
# CPU (install locally; run validate_harness.py before A100)
# ============================================================
regex>=2023.10.3 # \X grapheme cluster support — NOT stdlib re
Pillow>=10.0.0 # image rendering in datagen.py / textkit.py
numpy>=1.24.0 # ink-density measurement
datasets>=2.18.0 # corpus_prep.py (HF streaming)

# ============================================================
# A100 session (GPU inference)
# ============================================================

# PyTorch — DeepSeek-OCR-2 model card specifies torch==2.6.0
torch>=2.6.0

# Transformers — PINNED CAREFULLY:
# DeepSeek-OCR-2 model card specifies transformers==4.46.3
# Qwen3-VL requires transformers>=4.57 (not yet released — install from source)
#
# Strategy: use source install for Qwen3-VL; DeepSeek-OCR-2 has its own
# trust_remote_code model class so version matters less than for Qwen3-VL.
#
# Option A (recommended): install from source once, covers both models:
# pip install git+https://github.com/huggingface/transformers
#
# Option B: pin for DeepSeek-OCR only, accept that Qwen3-VL may need upgrade:
# transformers==4.46.3
#
# We leave this unversioned here so pip doesn't fight with the source install.
transformers

# Flash Attention — required by DeepSeek-OCR-2 (_attn_implementation="flash_attention_2")
# Build from source; binary wheels available for common CUDA versions:
# pip install flash-attn --no-build-isolation
# flash-attn>=2.7.3

accelerate>=0.28.0 # device_map="auto" support
sentencepiece>=0.2.0 # tokenizer dependency for several models
protobuf>=4.25.0

# ============================================================
# Optional / benchmark phase (P3)
# ============================================================
# PaddleOCR (if benchmarking PaddleOCR-VL):
# pip install paddlepaddle-gpu paddleocr
# Tesseract Python binding:
# pip install pytesseract (also install tesseract binary + tam traineddata)
# TEDS (table eval):
# pip install apted distance (for table tier metric)

YAML Metadata Warning:The task_ids "optical-character-recognition" is not in the official list: acceptability-classification, entity-linking-classification, fact-checking, intent-classification, language-identification, multi-class-classification, multi-label-classification, multi-input-text-classification, natural-language-inference, semantic-similarity-classification, sentiment-classification, topic-classification, semantic-similarity-scoring, sentiment-scoring, sentiment-analysis, hate-speech-detection, text-scoring, named-entity-recognition, part-of-speech, parsing, lemmatization, word-sense-disambiguation, coreference-resolution, extractive-qa, open-domain-qa, closed-domain-qa, news-articles-summarization, news-articles-headline-generation, dialogue-modeling, dialogue-generation, conversational, language-modeling, text-simplification, explanation-generation, abstractive-qa, open-domain-abstractive-qa, closed-domain-qa, open-book-qa, closed-book-qa, text2text-generation, slot-filling, masked-language-modeling, keyword-spotting, speaker-identification, audio-intent-classification, audio-emotion-recognition, audio-language-identification, multi-label-image-classification, multi-class-image-classification, face-detection, vehicle-detection, instance-segmentation, semantic-segmentation, panoptic-segmentation, image-captioning, image-inpainting, image-colorization, super-resolution, grasping, task-planning, tabular-multi-class-classification, tabular-multi-label-classification, tabular-single-column-regression, rdf-to-text, multiple-choice-qa, multiple-choice-coreference-resolution, document-retrieval, utterance-retrieval, entity-linking-retrieval, fact-checking-retrieval, univariate-time-series-forecasting, multivariate-time-series-forecasting, visual-question-answering, document-question-answering, pose-estimation

Tamil OCR Benchmark v1

The first independent, open benchmark of 2026-generation OCR-VLMs on Tamil, with structure-coverage-controlled evaluation and a grapheme-aware metric protocol.

Highlights

247-grapheme coverage matrix — full uyirmey grid + Grantha + split matras + Tamil numerals
Grapheme-cluster CER (primary metric) — edit distance over \X clusters after NFC normalisation, shown to be unbiased vs. codepoint-CER for Indic scripts
Compression × script-density study (Pillar 3) — Pillar 3 verdict: NO-GO
Document tiers v1: printed multi-column, tables/forms, Tanglish
Vision-only scrambled probe — removes decoder language prior to isolate visual confound in compression experiments

Gate results

Gate	Verdict
Pillar 3 (compression × density)	NO-GO
Base model selection	Qwen3-VL-2B

Dataset structure

data/
  images/<split>/<script>/<real|scrambled>/<idx>.png
  manifests/<split>.jsonl
    fields: id, split, script, mode, text, ground_truth,
            image_path, grapheme_count
results/
  gate_a_deepseek.jsonl    DeepSeek-OCR-2 predictions (all budgets)
  gate_b_qwen.jsonl        Qwen3-VL predictions (real mode)
  pillar3_verdict.json     Gate A decision
  base_select_verdict.json Gate B decision
eval/
  evaluate.py              grapheme_cer + codepoint_cer
  benchmark_spec.py        coverage matrix + metric definitions
  textkit.py               grapheme segmentation utilities

Metrics

Metric	Description
`grapheme_cer`	Edit distance over `\X` grapheme clusters / cluster count (primary)
`codepoint_cer`	Edit distance over NFC codepoints / codepoint count (Pillar 2 demo only)
`word_acc`	Fraction of reference words matched at same position
`teds`	Tree Edit Distance Score for table/form tier

Records

1200 image–text pairs across 3 scripts × 2 rendering modes (real / scrambled).

Citation

@misc{tamil-ocr-benchmark-2026,
  title  = {Tamil OCR Benchmark v1: Compression, Density, and Script},
  author = {Venkateswaran, Balaji},
  year   = {2026},
  url    = {https://huggingface.co/datasets/mvbalaji/tamil-ocr-benchmark}
}

License

Dataset: CC BY 4.0 Fonts: OFL 1.1 (Noto Sans family, Google) Code: Apache 2.0

Downloads last month: 852

Number of rows:

Total file size:

19.5 MB