HKU-Liangcai commited on
Commit
99288f3
·
verified ·
1 Parent(s): 8a73bb0

Define exact full-score Pass Rate and document payload integrity

Browse files
Files changed (1) hide show
  1. README.md +35 -0
README.md CHANGED
@@ -63,6 +63,41 @@ Verify a downloaded solve package with:
63
  python tools/verify_dataset.py
64
  ```
65
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
66
  ## Citation
67
 
68
  ```bibtex
 
63
  python tools/verify_dataset.py
64
  ```
65
 
66
+ ## Metrics
67
+
68
+ Official Pass Rate uses **exact full credit**: a task passes only when
69
+ `evaluation_complete == 1` and `task_score == 1.0`. Native per-task
70
+ thresholds and the raw verifier's `passed` flag do not determine this metric.
71
+ There is no rounding or tolerance: `0.999` is not a pass.
72
+
73
+ - **Pass Rate (%)** = 100 × number of completed full-score tasks / 97.
74
+ - **Score (0–100)** = 100 × sum of valid, completed `task_score` values / 97.
75
+ - Partial credit remains the native rubric score normalized to [0, 1].
76
+ Missing tasks, invalid scores, and incomplete evaluations contribute zero
77
+ to the fixed denominator and must be reported separately; incomplete runs
78
+ are not final benchmark results.
79
+ - Use one predeclared attempt per task. A subset may use its predeclared task
80
+ count as denominator, but must be labeled as a subset, not the 97-task result.
81
+
82
+ The runtime's `summary.json` records `metric_definition: exact-full-score`;
83
+ its `passed` column follows this rule and `native_passed` preserves the raw
84
+ verifier flag for diagnosis. Archived verifier outputs are unchanged. Historical
85
+ results using native thresholds or score tolerances must be recomputed from
86
+ raw scores before comparison. See the
87
+ [scoring guide](https://github.com/ApodexAI/FrontierAgent/blob/main/benchmarks/frontierchallenge/docs/scoring.md)
88
+ and [runtime update](https://github.com/ApodexAI/FrontierAgent/pull/30).
89
+
90
+ **Runtime rollout:** the runtime update is pending. Before reporting results,
91
+ check that your summary records `metric_definition: exact-full-score`;
92
+ older runtimes and Harbor's raw `passed` aggregate use the legacy rule.
93
+
94
+ ## Integrity
95
+
96
+ The root `README.md` is a mutable dataset card and is intentionally outside
97
+ `checksums.sha256`. Task files (including task-level READMEs), manifests,
98
+ registries, and verification tools remain checksummed. Payload changes require
99
+ regenerating their checksum entries; editing this card does not.
100
+
101
  ## Citation
102
 
103
  ```bibtex