anabogdan commited on
Commit
418a6cc
·
verified ·
1 Parent(s): 2461f50

Upload PROMPTS.md

Browse files
Files changed (1) hide show
  1. PROMPTS.md +122 -0
PROMPTS.md ADDED
@@ -0,0 +1,122 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # RoJBMO Prompt Templates
2
+
3
+ The exact prompt templates used in the RoJBMO evaluation pipeline: solution
4
+ generation, pointwise judging, and pairwise judging. Referenced from the
5
+ paper ("*RoJBMO: Evaluating Large Language Models on Junior Balkan
6
+ Mathematical Olympiad Problems*", SYNASC 2026), Section IV-C.
7
+
8
+ Placeholders in `{curly braces}` are filled in per problem at runtime; the
9
+ rest of the text is sent to the model unmodified.
10
+
11
+ ---
12
+
13
+ ## 1. Generation prompt
14
+
15
+ Elicits a candidate solution from each generation model.
16
+
17
+ ```text
18
+ You are an expert mathematician participating in the Junior Balkan Mathematical Olympiad (JBMO).
19
+
20
+ Solve the following {domain} problem completely, showing all steps clearly.
21
+ Write your solution in a clean, rigorous mathematical style.
22
+
23
+ PROBLEM ({year} · {domain}):
24
+ {statement}
25
+
26
+ Provide a complete, step-by-step solution:
27
+ ```
28
+
29
+ - `{domain}` — one of Algebra / Combinatorics / Geometry / Number Theory.
30
+ - `{year}` — the competition year of the source problem.
31
+ - `{statement}` — the problem statement, LaTeX, English.
32
+
33
+ Sampling: temperature 0.2, single user-turn message, no system prompt.
34
+
35
+ ---
36
+
37
+ ## 2. Pointwise judge prompt
38
+
39
+ Scores a single candidate solution against the reference solution across
40
+ five dimensions.
41
+
42
+ ```text
43
+ You are an expert JBMO (Junior Balkan Mathematical Olympiad) grader.
44
+ Evaluate the candidate solution below against the reference solution.
45
+
46
+ ═══ PROBLEM ({year} · {domain}) ═══
47
+ {statement}
48
+
49
+ ═══ REFERENCE SOLUTION ═══
50
+ {reference_solution}
51
+
52
+ ═══ CANDIDATE SOLUTION (by {model_label}) ═══
53
+ {candidate_solution}
54
+
55
+ ═══ SCORING INSTRUCTIONS ═══
56
+ Score each dimension 1–10. Be strict: a single unjustified logical step or gap in correctness
57
+ should cap that dimension at ≤ 6. Only award 9–10 when the solution is virtually flawless.
58
+
59
+ Dimensions:
60
+ • correctness (1–10): Every logical step valid, no hidden assumptions, conclusion fully justified.
61
+ • clarity (1–10): Notation clean, argument easy to follow, well-organised.
62
+ • completeness (1–10): All cases covered, no steps skipped, final answer clearly stated.
63
+ • elegance (1–10): Approach is clean, insightful, or avoids unnecessary computation.
64
+ • overall (1–10): Holistic quality — not necessarily the average of the above.
65
+
66
+ Respond ONLY with valid JSON (no markdown fences, no extra text):
67
+ {
68
+ "correctness": <int 1-10>,
69
+ "clarity": <int 1-10>,
70
+ "completeness": <int 1-10>,
71
+ "elegance": <int 1-10>,
72
+ "overall": <int 1-10>,
73
+ "reasoning": "<3-4 sentences: key strengths, key flaws, and whether the solution is correct>"
74
+ }/no_think
75
+ ```
76
+
77
+ - `{model_label}` — human-readable name of the model being judged (excluded
78
+ from the judge's own generation pool to avoid self-preference).
79
+ - The trailing `/no_think` suppresses extended-thinking preambles on models
80
+ that support the convention, so the response stays pure JSON.
81
+
82
+ Judge: GPT-4o, temperature 0.0, single user-turn message.
83
+
84
+ ---
85
+
86
+ ## 3. Pairwise judge prompt
87
+
88
+ Directly compares two candidate solutions on the same problem, using a
89
+ four-class verdict scheme. Each pair is judged in both orders (A/B and B/A)
90
+ to measure position bias.
91
+
92
+ ```text
93
+ You are an expert JBMO (Junior Balkan Mathematical Olympiad) grader.
94
+ Two candidate solutions are provided below. Determine whether each is mathematically correct.
95
+
96
+ ═══ PROBLEM ({year} · {domain}) ═══
97
+ {statement}
98
+
99
+ ═══ REFERENCE SOLUTION ═══
100
+ {reference_solution}
101
+
102
+ ═══ SOLUTION A (by {label_a}) ═══
103
+ {solution_a}
104
+
105
+ ═══ SOLUTION B (by {label_b}) ═══
106
+ {solution_b}
107
+
108
+ ═══ INSTRUCTIONS ═══
109
+ Classify into exactly one of four cases:
110
+ • "both_correct" — both solutions reach the correct answer by a valid mathematical argument
111
+ • "A_correct" — only Solution A is mathematically correct
112
+ • "B_correct" — only Solution B is mathematically correct
113
+ • "neither_correct" — neither solution is correct
114
+
115
+ Respond ONLY with valid JSON (no markdown fences, no extra text):
116
+ {
117
+ "verdict": "<both_correct|A_correct|B_correct|neither_correct>",
118
+ "reasoning": "<2-3 sentences explaining which solution is correct and why>"
119
+ }/no_think
120
+ ```
121
+
122
+ Judge: same as the pointwise prompt above.