Suxinchun/Jianghu108

English | 简体中文

Note: This is a machine translation. The Chinese version is authoritative. See the ModelScope link: https://www.modelscope.cn/datasets/suxinchun/Jianghu108

Updated: 2026-09-21

Jianghu108 is a multiple-choice test set of Chinese chess (xiangqi) jianghu endgame studies, provided in two versions — with and without board images — so that the spatial reasoning of both multimodal and text-only large language models can be measured. All 108 problems are jianghu endgame studies, some of them ancient and some newly composed. Each problem contains the following fields: the problem name, the source (in every case a link to an explanation video by the Bilibili creator 从宽象棋2023; copyright remains with the original author, this test set does not include any video content and all images are rendered by ourselves), the requirement (Red wins / draw), the coordinates of the Red pieces and of the Black pieces (all coordinates are written as Chinese piece character–letter–number), five options A–E (the fifth is always "none of the above"), and the index of the correct answer.

Translation conventions

• Pieces are not translated on the board images. The board images are shared by the Chinese and English versions: the pieces are drawn as Chinese characters exactly as in the data (Red 帥 仕 相 馬 炮 兵 車, Black 将 士 象 馬 砲 卒 車), and the river is still written 楚 河 / 汉 界.

• In the English text, each piece character has a fixed English name. Where Chinese uses two different characters, English uses two different names:

  Character	English
  帥	General
  将	King
  仕	Advisor
  士	Guard
  相	Elephant
  象	Officer
  馬	Horse
  車	Chariot
  炮	Cannon
  砲	Catapult
  兵	Soldier
  卒	Pawn

  (帥/仕/相/炮/兵 are the Red characters, 将/士/象/砲/卒 the Black ones; 馬 and 車 are shared by both sides.)

• Move notation in this English document is written as <English piece name> <origin> <target>, e.g. Chariot B1 E1, Cannon B7 B10+, Soldier F9 x F10+. x means the move captures the piece on the target square, + means the move gives check. The JSON file itself keeps the Chinese notation (車B1 E1); the two are equivalent.

• The problem names are kept in Chinese, with their pinyin given in section 2.5.

• Jianghu (江湖) endgame studies are the street-corner endgame compositions of Chinese chess, typically rich in traps and in "counter-check" resources.

---

1. Introduction: why jianghu endgame studies are well suited to evaluating large language models

1. The rules are clear. Anyone who knows Chinese chess knows them; there are just a few and no more. Some jianghu endgame studies do involve disputes about rules (especially the difference between the Asian rules and the Chinese rules on perpetual check and perpetual chase), but such problems are excluded from this set: only problems whose main line involves no rule dispute are included, so there is no room to quibble about the rules. Of course, either side could deliberately play abnormally — refusing a mate or chasing perpetually — but that is outside the scope of consideration, since we require optimal play from both sides.
2. There is an image, and the image can be described exactly in words. A position can be given either as a board image or losslessly as text ("piece + coordinate"), so text-only and multimodal models can both use the same problem. The prompt is identical whether or not an image is attached.
3. Jianghu endgame studies always contain tempting moves; they look easy, and large language models take the bait just like humans do. But in fact there is the exquisite move that answers a mate with a mate, which makes them really hard. This is the most elegant part.

For example, problem 23, 连火阵 (Lián Huǒ Zhèn): it looks like the classic sacrifice of both Chariots to bring the Cannon down — give up the two Chariots on the H file, bring the Cannon down with check and win. How bold, how plausible a solution. But after Red's 車H3 D3+, Black takes with a Pawn (卒C3 x D3); Red sacrifices again with 車H4 D4+, Black takes with the Horse (馬E2 x D4), and the Black Cannon is uncovered; Red then plays 炮H2 H8+, and Black answers exactly with E1 E8+, answering the check and mating Red on the spot. The correct first move is a quiet one, 兵E9 D9, after which the two sides struggle on and the game ends in a draw. It is very hard to find.

4. It fits a multiple-choice format. The requirement is Red wins or draw, and the question is what the first move is, with four options. Some positions do not fit this format — for instance a Red loss, or a position where the first move has no choice at all (a single check that must be played, since otherwise Red is mated). Those are not included; this set only contains problems that fit, that is, where the first move has more than one candidate.
5. The positions look very similar to one another, so even a leaked problem cannot be memorized. Moreover, there are hardly any standard textual descriptions of Chinese chess endgame studies — almost all of them are videos; unless the model is told the name of the study, it has absolutely no way to find the problem.
6. Anyone who can play chess can read them. These are not problems that deliberately split hairs. Everyone can see plenty of problems of that kind.
7. This is completely new content. As far as the author knows, nobody has ever made a test set of Chinese chess jianghu endgame studies before.

---

2. Contents and format

2.1 File list

File / directory	Description
江湖残局.json	The data set itself, 108 problems
boards/board_001.png … board_108.png	One board image per problem, 792×870
images/	English statistics charts and word clouds (used by this English page)
各题明细.md	Per-problem details: statement, board image, four options and the correct answer
各题名拼音.txt	The 108 problem names: id, Chinese name, pinyin
test_deepseek_xiangqi.py	Evaluation script
检查笔误.mjs	Data validation script (node 检查笔误.mjs)

2.2 Prompt template

The header field is identical for all 108 problems. Its English rendering is:

This is a high-difficulty Chinese chess (xiangqi) jianghu endgame study. The rules follow the Asian rules: perpetual check and perpetual chase are forbidden, everything else is unrestricted. There is no rule dispute in this problem, that is, the correct answer is not obtained through a rule dispute such as perpetual check. The rule that the game is drawn when neither side captures a piece for 60 moves is also disregarded. Red moves first. The board is the standard 10-row, 9-column board, Black at the top and Red at the bottom. The intersection of the left-hand edge and the bottom edge from Red's point of view is the lower-left point of the board, i.e. the origin; starting from the origin, the columns from left to right are columns A–I (9 columns) and the rows from bottom to top are rows 1–10 (10 rows). The pieces and moves of both sides are expressed in this coordinate system. A move is recorded as piece, origin square, target square; a "+" marks that the move gives check, an "x" marks that the move captures the piece on the target square, and no mark means that the move neither captures a piece nor gives check. The initial position is guaranteed to be legal. Given the current position, which first move by Red can be the correct answer? Put the letter of your answer between <box> and </box>, e.g. <box>X</box>.

The prompt declares two rule conventions that affect evaluation: perpetual check and perpetual chase are forbidden (the correct answer does not depend on a rule dispute), and the "draw after 60 moves without a capture" rule is disregarded, as it was disregarded in the original videos. The required result of each problem (Red wins / draw) is not written into header; it is given separately by the circumstance field of each problem and inserted when the prompt is assembled (see section 4).

2.3 Data structure

The top level is {"problems": [ ... ]}, and the field order of each problem is:

id, name, source, header, circumstance, Red, Black, options, correctAnswer

Field	Description
id	Problem number, 1–108
name	Name of the study (e.g. 双蛟翦水, 长坂坡, 螳臂挡車), in Chinese. All 108 names are distinct; please be careful not to give the name to the model during evaluation!
source	Link to the Bilibili video for this problem; as above, do not give it to the model
header	The statement; identical for all 108 problems, see 2.2
circumstance	The result of this problem; only two values occur: 在双方最优应对下，本局红方胜 (with optimal play by both sides, Red wins) and 在双方最优应对下，本局红方无法取胜，为和棋 (with optimal play by both sides, Red cannot win; the game is a draw). There is no "Red loses"
Red / Black	Positions of the Red and Black pieces, e.g. 車B1 炮B4 炮B7 帥F2 兵F9 (i.e. Chariot B1, Cannon B4, Cannon B7, General F2, Soldier F9)
options	Five options; the first four are concrete moves (e.g. A：炮B7 B10+, i.e. Cannon B7 B10+), the fifth is always E：以上选项都不对 (none of the above)
correctAnswer	Letter of the correct option (A–E)

2.4 Coordinate system and notation

• The board has 10 rows and 9 columns, Black at the top, Red at the bottom.
• The origin is the intersection of the left-hand edge (from Red's point of view) with the bottom edge (Red's own baseline): from left to right the columns are A–I (9 columns), from bottom to top the rows are 1–10 (10 rows). This is done to avoid the problems caused by the mirror symmetry of the traditional Chinese chess notation between Red and Black, which would cost a large language model a great deal of effort; the traditional notation is therefore not used.
• Red's baseline is row 1 and Black's baseline is row 10; the river lies between row 5 and row 6, so Red's half is rows 1–5 and Black's half is rows 6–10.
• A move is written as "piece, origin, target", e.g. 炮B7 B10+ — in English, Cannon B7 B10+. + means the move gives check, x means the move captures the piece on the target square; when neither is marked, the move neither captures nor gives check.

2.5 Piece characters, and the names of the problems

Throughout the data set the characters are fixed (the board and the option texts agree): Red uses 帥 仕 相 馬 炮 兵 車 and Black uses 将 士 象 馬 砲 卒 車. That is, "Horse" is written 馬 for both sides; "Cannon" is 炮 for Red and 砲 for Black; and "Advisor/Guard" and "Elephant/Officer" are split between Red and Black. The English names are given in the table under "Translation conventions" above.

The 108 names are all in Chinese; their pinyin is given in 各题名拼音.txt (id, Chinese name, pinyin).

2.6 Data validation

Every problem is validated by 检查笔误.mjs, which checks:

• no square is occupied twice; each side has exactly one General / King;
• Red uses only Red characters and Black only Black characters (for instance 士 and 象 may not appear on the Red side);
• at the start neither side is in check, and the two Generals do not face each other;
• the General/King and the Advisors/Guards are inside the palace; the Elephants/Officers are on the 7 elephant points of their own half (Red: A3 C1 C5 E3 G1 G5 I3; Black: A8 C6 C10 E8 G6 G10 I8);
• every Soldier/Pawn could have reached its square from its starting position (forward only, and sideways only after crossing the river);
• all four options are legal moves in the given position: they follow each piece's movement rules (including the horse's blocked leg, the elephant's blocked eye and the cannon's capture over exactly one screen), and after the move the moving side is not in check and the two Generals do not face each other;
• a move marked x really does capture a piece on the target square, and that piece is not a General or King; a move marked + really does give check; the target square may not hold a piece of one's own side;
• no side exceeds the piece limits: at most 2 Chariots, 2 Horses, 2 Cannons, 2 Advisors/Guards, 2 Elephants/Officers, 5 Soldiers/Pawns and 1 General/King;
• circumstance is one of the two standard strings, and the spacing and the x / + marks are formatted consistently;
• the correct option is itself a legal move.

These positions all come from explanations by @从宽象棋2023; some of them may be explained incorrectly there, so criticism and corrections are welcome.

---

3. Statistics

3.1 Overview

Item	Value
Number of problems	108
Result	Red wins 46, draw 62
Distribution of the correct option	A 22, B 22, C 22, D 22, E 20 (E = "none of the above")
Red piece count	2–12 pieces, mean 6.21
Black piece count	3–16 pieces, mean 7.08
Both sides together	7–23 pieces, mean 13.30
Nature of the 432 options	75 captures, 268 giving check
Nature of the 108 correct answers	13 captures, 61 giving check
Names	all 108 distinct; 8 names of 3 characters, 96 of 4, 3 of 5, 1 of 8
First moves	the first moves of the four options (origin → target) take 140 distinct values, those of the correct answers 66; all moves of the four options (piece + origin → target) take 307 distinct values, those of the correct answers 77

3.2 Distribution of the correct option

The distribution of the correct answer over the five options A–E; the dashed line is the perfectly uniform value of 21.6.

3.3 Distribution of results

The 108 problems consist of 46 Red wins and 62 draws.

3.4 Names

The distribution of the lengths of the 108 names: 3, 4, 5 and 8 characters.

The word cloud built from the 108 full names; since each name occurs once, no name is larger than another.

The word cloud counted by character; the font size is the number of times the character occurs in the 108 names.

3.5 Number of pieces

The distribution of the number of Red pieces per problem (2–12, mean 6.21).

The distribution of the number of Black pieces per problem (3–16, mean 7.08).

The average number of each Red piece (Chariot, Horse, Cannon, Soldier, Advisor, Elephant) and its ratio to the limit for that piece.

The average number of each Black piece (Chariot, Horse, Catapult, Pawn, Guard, Officer) and its ratio to the limit for that piece.

3.6 Position distribution of the pieces

Each chart uses the board as its background (the same one as in boards/: double outer frame, river, 楚河/汉界, the cannon points, and the coordinates A–I and 1–10 outside the board). The number in each circle is the total number of times a piece of that side occurred on that square over all problems.

The position distribution of all pieces of both sides over the 108 problems.

The position distribution of the Red pieces.

The position distribution of the Black pieces.

The position distribution of the Red Chariot.

The position distribution of the Red Horse.

The position distribution of the Red Cannon.

The position distribution of the Red Soldier.

The position distribution of the Red Advisor.

The position distribution of the Red Elephant.

The position distribution of the Red General.

The position distribution of the Black Chariot.

The position distribution of the Black Horse.

The position distribution of the Black Catapult.

The position distribution of the Black Pawn.

The position distribution of the Black Guard.

The position distribution of the Black Officer.

The position distribution of the Black King.

3.7 Distribution of first moves

The first moves among the four options, grouped by piece; the 10 most frequent.

The first moves of the 108 correct answers, grouped by piece; the 10 most frequent.

3.8 Nature of the moves

The number of capturing moves among the 432 candidates (75).

The number of capturing moves among the 108 correct answers (13).

The number of moves giving check among the 432 candidates (268).

The number of moves giving check among the 108 correct answers (61).

3.9 Whether Black already has a mate in one

Of the 108 problems, the number in which Black can already mate Red in one move at the start: 53, and the number in which Black cannot: 55.

4. Prompt construction

4.1 Assembly

The prompt sent to the model consists of four parts:

1. the part of header up to "The initial position is guaranteed to be legal.";
2. the conclusion sentence derived from that problem's circumstance: for a win, "with optimal play by both sides, Red wins"; for a draw, "with optimal play by both sides, Red cannot win, and the game is a draw";
3. the question part of header: "Given the current position, which first move by Red can be the correct answer? Put the letter of your answer between and , e.g. X";
4. three lines: "Red pieces: …", "Black pieces: …" and "Options:" followed by the five options, one per line.

The prompt actually sent to the models is the original Chinese one; the English above is a translation.

4.2 The complete prompt for problem 1

这是一局中国象棋高难度的江湖残局，规则采用亚洲棋规，禁止长将长捉，其他不限，本局没有规则争议，也就是正解不依靠一将一杀等规则争议取得。此外也不考虑双方各60步内均不吃子则和棋的规则。红先。标准的10行9列棋盘，黑上红下，以红方视角的左侧和红方视角的底线交叉点为棋盘左下点，即原点，从原点出发从左到右的列记为A-I列共9列，从下到上的行记为1-10行共10行。双方的棋子和走法均在该坐标系下表示。棋子走法记录为棋子、原位置、目标位置，标“+”表示走子后产生将军，标“x”表示吃目标位置上的子，没有标注则代表不吃子且不将军。保证初始局面为合法局面。已知在双方最优应对下，本局红方胜，那么，在当前局面下，本局能成为正解的红方第一步是？请将你的答案选项字母放在<box>与</box>之间，如<box>X</box>
红方棋子位置: 車B1 炮B4 炮B7 帥F2 兵F9
黑方棋子位置: 馬A1 卒A2 卒A3 砲A5 士D10 車E5 将E10 士F10 象G6 卒I2 砲I5
选项：
A：炮B7 B10+
B：炮B4 E4+
C：車B1 E1
D：兵F9 x F10+
E：以上选项都不对

In English notation (equivalent, for reference):

Red pieces:   Chariot B1, Cannon B4, Cannon B7, General F2, Soldier F9
Black pieces: Horse A1, Pawn A2, Pawn A3, Catapult A5, Guard D10, Chariot E5, King E10, Guard F10, Officer G6, Pawn I2, Catapult I5
Options:
A: Cannon B7 B10+
B: Cannon B4 E4+
C: Chariot B1 E1
D: Soldier F9 x F10+
E: none of the above

4.3 Multimodal input

In the vision version the board image boards/board_XXX.png of that problem is appended after the text above, with no explanatory text at all, to keep the comparison fair. The board image contains only the board, the pieces, the coordinates outside the board, and the move markers of the four options: four curved arrows in different colours for A/B/C/D (red, blue, green, purple), each with a letter of the same colour next to its starting point; the tip of a moving arrow lands exactly on the target point, while the tip of a capturing arrow stops at the edge of the captured piece; the captured piece is drawn as usual. There is no title, no piece list and no option text on the image; the river is marked 楚河/汉界 in Chinese, and the coordinate letters and numbers are present.

4.4 Answer extraction and judging

• Correct answer: the letter (A–E) inside the last <box>X</box> of the reply is compared with correctAnswer; if the reply does not follow the format, the last isolated A–E letter in the text is used instead.
• An empty reply does not count as correct.

---

5. Test script and results

test_deepseek_xiangqi.py: --model flash|pro|vision; the three models share the same prompt. Arguments:

Argument	Effect
--model flash|pro|vision	choose the model; vision additionally attaches the board image
--ids 1,5,23	test only the given problem numbers
--batches N	number of concurrent requests, default 108 (one request per problem)
--think / --no-think	thinking mode (default on) / answer directly
--max-tokens N	output limit in thinking mode, default 131072
--detail low|high|original|auto	vision image detail
--dry-run	print the assembled prompt without making requests
--base-url / --api-key	API base URL and key

---

6. See also

• 痛快80 (Tsūkai 80): a Go life-and-death (tsumego) problem set, 80 problems from Maeda Nobuaki's Tsūkai, presented as multiple-choice questions.
• 三国杀数据集 (Sanguosha dataset): a Sanguosha (Legends of the Three Kingdoms) dataset.
• AI大宝 (AI Dabao): a Sanguosha AI studio.