savitr-bf16

savitr-bf16 — full-precision terse electoral-roll OCR (distilled Surya)

Full-precision (bf16) merged weights for savitr's distilled terse model. This is the checkpoint that gojiberries/savitr (MLX 8-bit) was quantized from — same model, same training run, higher precision.

Use gojiberries/savitr if you just want to read electoral rolls on Apple Silicon: it is smaller (~800 MB vs 1.33 GB), faster, and what pip install savitr fetches automatically. Use this repo if you want to:

• re-quantize to a different format or bit width (MLX 4/8-bit, GGUF, AWQ)
• continue training or distilling from the merged weights
• evaluate without quantization error

Base model = datalab-to/surya-ocr-2 (650M Qwen3.5-VL-style OCR), self-distilled to emit one compact pipe-delimited line per voter instead of verbose HTML:

epic|name|relation(F/H/M)|relative_name|house|age|sex

~5× fewer decode tokens than the HTML output, at the teacher's accuracy.

Usage

Re-quantize to MLX (what the published 8-bit is):

python -m mlx_vlm convert --hf-path gojiberries/savitr-bf16 \
    --mlx-path savitr-mlx-8bit -q --q-bits 8

Load directly:

from transformers import AutoModelForImageTextToText, AutoProcessor
model = AutoModelForImageTextToText.from_pretrained("gojiberries/savitr-bf16")
proc = AutoProcessor.from_pretrained("gojiberries/savitr-bf16")

Parse the output with savitr's value-anchored parser, which stays column-aligned even when the model drops a field:

from savitr.rolls.parse import parse_terse, TERSE_PROMPT

How it was trained

Teacher = full Surya (surya-ocr-2) OCRs roll pages to HTML; a parser cleans them into terse targets; the model is LoRA-fine-tuned on (page image → terse rows) — 450 pages drawn from constituencies held out of the eval, 1 epoch, for $0 on a free Kaggle T4 — then the adapter is merged into the base weights. The terse format is the only behavioral change; reading ability is inherited from Surya.

Evaluation (out-of-sample, vs the Surya teacher)

Held-out constituencies never seen in training (37 pages, 1,076 teacher voters):

Per-voter record similarity 98.7%, whole-page similarity 92.9% (1 − normalized edit distance). Fidelity = agreement with the teacher's output; absolute accuracy ≈ these × Surya's own ~93–95%. Measured on the 8-bit quantization of these weights, so bf16 is a floor, not a ceiling.

Limitations

v0.2 (450 training pages, AC-holdout). Trained on Manipur 2025 English rolls; other states and scripts are out of distribution. The tokenizer, processor, and chat template here come from the base model — the training run's own tokenizer was written by a newer transformers and is not loadable by current mlx-vlm.

License & attribution

Derived from datalab-to/surya-ocr-2; its license governs use of these weights. savitr's code is MIT (github.com/in-rolls/savitr). Electoral rolls are public records published by the Election Commission of India.