Modelpublic

kailasa-ngpt/gemma-4-31b-tamil-devanagari-ocr-bf16

sourceHugging Faceapache-2.0updated 29d agoView on Hugging Face

0likes48downloads

Model Card

Gemma 4 31B Tamil + Devanagari OCR — BF16

This is the primary-quality merged checkpoint produced by the workspace's Tamil and Devanagari OCR fine-tune. It starts from google/gemma-4-31B-it and merges the selected rank-32 language-layer LoRA into Google's original BF16 weights. The vision tower stayed frozen. Sanskrit is covered where it is written in Devanagari; this model has not been validated for other Sanskrit scripts such as Grantha.

Training

—Distillation dataset: kailasa-ngpt/gemini-3.7-flash-ocr-26-aug-2026
—Target rows: 8,374 Devanagari and 4,066 Tamil; 68 Telugu rows excluded
—Balanced training stream: 16,748 examples through deterministic Tamil oversampling
—Optimized examples: 4,000 (1,000 steps, physical batch 4)
—QLoRA: NF4 load, rank/alpha 32/32, 410 language-layer targets
—Learning rate: 1e-4 cosine; best checkpoint at step 950
—Best clean-book validation loss: 0.407109946
—Sequence/image budgets: 8,192 total tokens and 1,120 vision tokens
—Training peak: 31.13 GiB allocated on one RTX PRO 6000 96GB

Evaluation

All text metrics preserve raw model Unicode and line boundaries. The primary metric is extended-grapheme-cluster CER (gCER).

Gate	gCER mean	gCER median	gCER clipped at 100%	Catastrophic	4,096-token caps
Independent 80-page gate	25.68%	11.12%	17.42%	7/80	4/80
Untouched base, same pages	374.74%	17.56%	25.99%	11/80	6/80

The fine-tune beat the base on 66/80 paired pages. A 20,000-replicate, script-stratified bootstrap estimated 99.505% probability that its raw mean gCER is lower; the very wide interval is caused by unbounded CER on base-model repetition loops.

On the untouched public benchmark's 350 natural-text cases, this BF16 artifact scored 3.238% gCER (Tamil 5.512%, Devanagari 0.965%) and 53.14% exact match. Its aggregate gCER difference from the base was not statistically resolved on that short-text benchmark.

vLLM

From the workspace root:

bash

scripts/serve_vllm.sh artifacts/gemma4-31b-tamil-devanagari-ocr-bf16
.venv-serve/bin/python scripts/ocr_vllm_client.py page.png --output page.txt

The client uses the exact training prompt, temperature-zero decoding, preserves output verbatim, and cancels long exact grapheme repetition cycles. Exit code 2 means the token limit was reached; exit code 3 means the repetition guard fired. The artifact pins the 1,120-token image budget in both model and processor configuration, so ordinary vLLM loading matches the fine-tuning input budget.

Limitations and terms

The source dataset card declares license other, and its labels are Gemini teacher outputs rather than human ground truth. It also documents teacher-side normalization. Review the source-corpus rights, dataset terms, Gemini terms, and the base model's Apache 2.0 license before redistribution or commercial deployment. OCR errors, including occasional autoregressive loops, remain; do not treat the output as a faithful scholarly edition without review.