Per-table card variant · Experimental
Micro card (≤1K tokens)
Smallest viable card: caption + inline Markdown table + 6-line YAML. Targets the 4K-context open-model tier.
In one paragraph
Stripped to absolute minimum: caption + PDF page + the Markdown table inline + a 6-line YAML frontmatter (variant tag, table number, source SHA, faithfulness level). Roughly 1 KB / 250 tokens per card. Designed for open models with strict 4K context windows.
How the inputs are generated
Generation · 01Generator script
evaluation_runs/generate_card_variants.py:render_micro_1kInput sources
- • pipeline-v0.6.1 cards
- • Docling-extracted Markdown tables
AI use
No — pure deterministic transformation
OCR / re-OCR
Inherits from the upstream pipeline variant
Approximate processing time
~2 seconds for all 407 cards.
Resource intensity
Low — CPU-only post-processing, runs in seconds
Determinism
Deterministic (same input → same output, byte-identical)
Output location
card_sets/pipeline-v0.7-micro-1k/Cards produced
407 cards
Introduced
v0.7 variant family, 2026-05-22.
Evaluation results
Diagnostic · 02Avg open-tier pass rate
40%
Typical card size
~1.2 KB per card
Evaluation cycle
Cycle 14
Relative to v0.6.1 baseline
+13 percentage points open-tier versus v0.6.1 baseline
Caveats and known limitations
Scope · 05- • Strips all surrounding context — questions requiring methodology (e.g. interpolation) lose the explanatory paragraphs.
- • Markdown table format performs worse than CSV on this benchmark; csv-only is the preferred compact format.