seat_me_wind string | seat_me_points int64 | seat_right_wind string | seat_right_points int64 | seat_left_wind string | seat_left_points int64 | seat_top_wind string | seat_top_points int64 | dora_indicator string | dora_value string | round_wind string | round_num int64 | tiles_left int64 | room string | hand list | disc_right list | disc_left list | disc_top list | call_me dict | call_right dict | call_left dict | call_top dict |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
north | 25,000 | east | 25,000 | west | 25,000 | south | 25,000 | 1s | 2s | east | 1 | 65 | Friendly Match - 4-Player South | [
"2m",
"1p",
"3p",
"4p",
"8p",
"4s",
"6s",
"east",
"west",
"west",
"north",
"R",
"0s"
] | [
"1p"
] | [
"1p"
] | [
"W",
"1p"
] | {
"actions": [],
"tiles": []
} | {
"actions": [],
"tiles": []
} | {
"actions": [],
"tiles": []
} | {
"actions": [
"pon"
],
"tiles": [
"south"
]
} |
west | 25,000 | north | 25,000 | south | 25,000 | east | 25,000 | G | R | east | 1 | 67 | Friendly Match - 4-Player South | [
"6m",
"7m",
"2p",
"3p",
"1s",
"1s",
"2s",
"2s",
"3s",
"7s",
"9s",
"south",
"W",
"W"
] | [] | [
"W"
] | [
"north"
] | {
"actions": [],
"tiles": []
} | {
"actions": [],
"tiles": []
} | {
"actions": [],
"tiles": []
} | {
"actions": [],
"tiles": []
} |
east | 25,000 | south | 25,000 | north | 25,000 | west | 25,000 | 9s | 1s | east | 1 | 69 | Friendly Match - 4-Player South | [
"5m",
"7m",
"7m",
"8m",
"4p",
"7p",
"7p",
"1s",
"4s",
"6s",
"7s",
"west",
"west",
"west"
] | [] | [] | [] | {
"actions": [],
"tiles": []
} | {
"actions": [],
"tiles": []
} | {
"actions": [],
"tiles": []
} | {
"actions": [],
"tiles": []
} |
east | 25,000 | south | 25,000 | north | 25,000 | west | 25,000 | 6p | 7p | east | 1 | 69 | Friendly Match - 4-Player South | [
"3m",
"9m",
"9m",
"1p",
"2p",
"4p",
"5p",
"8p",
"9s",
"south",
"south",
"west",
"north",
"R"
] | [] | [] | [] | {
"actions": [],
"tiles": []
} | {
"actions": [],
"tiles": []
} | {
"actions": [],
"tiles": []
} | {
"actions": [],
"tiles": []
} |
north | 25,000 | east | 25,000 | west | 25,000 | south | 25,000 | 0p | 6p | east | 1 | 66 | Friendly Match - 4-Player South | [
"0m",
"7p",
"9p",
"1s",
"4s",
"4s",
"7s",
"8s",
"9s",
"east",
"north",
"north",
"W",
"8p"
] | [
"R"
] | [
"G"
] | [
"W"
] | {
"actions": [],
"tiles": []
} | {
"actions": [],
"tiles": []
} | {
"actions": [],
"tiles": []
} | {
"actions": [],
"tiles": []
} |
south | 25,000 | west | 25,000 | east | 25,000 | north | 25,000 | south | west | east | 1 | 68 | Friendly Match - 4-Player South | [
"1m",
"8m",
"2p",
"5p",
"1s",
"1s",
"3s",
"3s",
"7s",
"8s",
"9s",
"west",
"G"
] | [] | [
"9p"
] | [] | {
"actions": [],
"tiles": []
} | {
"actions": [],
"tiles": []
} | {
"actions": [],
"tiles": []
} | {
"actions": [],
"tiles": []
} |
SightRead — Mahjong pack v0.1 (ground truth + prompt)
Hand-verified ground truth for the SightRead benchmark: exact game-state extraction from screenshots by vision-language models.
Not whether a model can play mahjong — whether it can see the board.
- Repository: https://github.com/AnsteinHuynh/sightread-bench
- Package:
pip install sightread-bench
What's in here
| File | Purpose |
|---|---|
pack.yaml |
field spec: 22 fields with comparator types + grounding traps |
prompt.md |
the extraction prompt shown to every model |
gt/*.json |
6 hand-verified full game states (sym-1..3, ind-1..3) |
Each GT state contains: 4 seat winds + 4 seat point counts, dora indicator and computed dora value, round wind/number, wall tile count, room text, the complete 14-tile hand (canonical multiset), three ordered discard rivers, and four pon/chii/kan call states.
What's deliberately NOT here
The screenshots. Game captures are copyrighted by their platforms
(MahjongSoul), so this repo ships ground truth + prompt only. To run the
benchmark, capture your own screenshots of the corresponding states and drop
them in packs/mahjong/images/ (gitignored by design) — the GT filenames
name which state each image must show.
Grounding trap
prompt.md contains an example hand that matches NO state
(1m 2m 3m 1p 2p 3p 1s 2s 3s N S W N R). A model whose extracted hand equals
it echoed the prompt instead of reading the image — SightRead's
grounding-fidelity signal, declared in pack.yaml and flagged
automatically by the scorer.
Scoring
Format-agnostic: responses are parsed line-wise with a fallback pass that
recovers fields anywhere in the text. Hands score as unordered multisets
(exact + per-tile F1); discard rivers as ordered sequences; numbers exact;
text dash/case/quote-insensitive. Full logic in
src/sightread/ at the repository above.
Citation
@software{sightread2026,
title = {SightRead: Exact Game-State Extraction Benchmark for Vision-Language Models},
author = {AnsteinHuynh},
year = {2026},
url = {https://github.com/AnsteinHuynh/sightread-bench}
}
- Downloads last month
- 36