H

HTR · VLM as Annotator

LLM as annotator · OpenRouter
·
sandbox · lang
🖼

No page selected

Use Upload or pick a sandbox in the sidebar.

Settings

Language / script

Substituted as {language} in every system prompt (OCR + Expert + Judge).

Transcription policy (= guidelines)

Source of truth for HOW to transcribe (preserve / expand / modernise). Substituted as {guidelines} in OCR + Expert + Judge system prompts.

Temperature (all models)
Output format
JSON template the model must fill

The model is asked to return this exact shape, with placeholder values replaced by content from the page. Use null for fields that have no value on the page; arrays receive one entry per line/item. Try a preset above to see common shapes, then edit to match your needs.

OCR model (OpenRouter slug) browse models ↗
N in-context examples
OCR system prompt

{language} + {guidelines} substituted at call-time. Empty = the default for the selected Output mode above.

OCR user prompt (zero-shot)

Empty = use the default shown as placeholder above.

browse OpenRouter models ↗
MoE expert A
MoE expert B
MoE judge (invoked only if A and B disagree)
Expert system prompt
Expert user prompt

{ocr_text} = OCR prediction. Empty = default (placeholder).

Judge system prompt
Judge user prompt

Variables: {ocr_text}, {expert_a_text}, {expert_b_text}, {expert_a_conf}, {expert_b_conf}, {expert_a_corrections}, {expert_b_corrections}. Empty = default.

OpenRouter API key

Sent in the X-API-Key header on each request. Stored only in localStorage of this browser, never on the server.

An OPENROUTER_API_KEY env variable is detected on the server and will be used as fallback.

ICL pool ·

⇩ Export JSONL

  • No ICL for this page yet.

    This sandbox is a 0-shot demo — or all its examples were deleted.

    Validate a correction or paste a transcription with the quick-add tool.

OCR prompt preview

Loading…

Session cost

Reported by OpenRouter via usage.include = true on each request.

Total cost
Requests
Input tokens
Output tokens

Some models don't report a per-call usage.cost. In that case tokens are counted but the cost stays at $0 — fall back to the OpenRouter activity dashboard for accurate billing.

Cite this work

If you use HTR · VLM as Annotator in your research, please cite it as:

APA

BibTeX 

      

      
The MoE post-correction strategy is adapted from internal pipeline scripts; please also acknowledge the underlying VLM providers (OpenRouter routes Anthropic, Google, OpenAI, Mistral, etc.).