Skip to content

Commit 58f505a

Browse files
committed
feat: add data/community runs
1 parent 8792aa0 commit 58f505a

4 files changed

Lines changed: 98 additions & 0 deletions

File tree

Lines changed: 13 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,13 @@
1+
{
2+
"model": "cerebras/gpt-oss-120b",
3+
"strategy": "default",
4+
"deck": "Red Deck",
5+
"stake": 1,
6+
"seed": "OOOO155",
7+
"challenge": null,
8+
"version": "0.3.0",
9+
"name": "Unknown Name",
10+
"description": "Unknown Description",
11+
"author": "BalatroBench",
12+
"tags": []
13+
}
Lines changed: 36 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,36 @@
1+
{
2+
"started_at": "2025-08-26T18:05:21.496076",
3+
"completed_successfully": true,
4+
"total_decisions": 23,
5+
"run_duration_seconds": 86.90119,
6+
"completed_at": "2025-08-26T18:06:48.397266",
7+
"total_responses": 23,
8+
"run_won": false,
9+
"final_score": 0,
10+
"final_money": 7,
11+
"ante_reached": 2,
12+
"final_round": 3,
13+
"blinds_defeated": 2,
14+
"boss_blinds_defeated": 0,
15+
"decisions_per_minute": 16.677341776741148,
16+
"score_per_decision": 0.0,
17+
"money_per_decision": 0.30434782608695654,
18+
"average_response_time_ms": 0.0,
19+
"jokers_acquired": 1,
20+
"consumables_used": 1,
21+
"shop_visits": 21,
22+
"rerolls_used": 0,
23+
"total_money_spent": 9,
24+
"peak_money_reached": 10,
25+
"hands_played_total": 13,
26+
"discards_used_total": 0,
27+
"money_efficiency_ratio": 0.7,
28+
"resource_management_score": 0.3,
29+
"hand_type_distribution": {},
30+
"decision_consistency_score": 0.7798652933595939,
31+
"risk_assessment_score": 0.6678367443917438,
32+
"joker_synergy_score": 0.0,
33+
"failed_requests": 0,
34+
"total_reasoning_length": 5113,
35+
"average_reasoning_complexity": 222.30434782608697
36+
}
Lines changed: 13 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,13 @@
1+
{
2+
"model": "cerebras/qwen-3-235b-a22b-thinking-2507",
3+
"strategy": "default",
4+
"deck": "Red Deck",
5+
"stake": 1,
6+
"seed": "OOOO155",
7+
"challenge": null,
8+
"version": "0.3.0",
9+
"name": "Unknown Name",
10+
"description": "Unknown Description",
11+
"author": "BalatroBench",
12+
"tags": []
13+
}
Lines changed: 36 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,36 @@
1+
{
2+
"started_at": "2025-08-26T18:54:32.051893",
3+
"completed_successfully": true,
4+
"total_decisions": 15,
5+
"run_duration_seconds": 101.404178,
6+
"completed_at": "2025-08-26T18:56:13.456071",
7+
"total_responses": 15,
8+
"run_won": false,
9+
"final_score": 0,
10+
"final_money": 2,
11+
"ante_reached": 1,
12+
"final_round": 2,
13+
"blinds_defeated": 1,
14+
"boss_blinds_defeated": 0,
15+
"decisions_per_minute": 9.604712712370869,
16+
"score_per_decision": 0.0,
17+
"money_per_decision": 0.13333333333333333,
18+
"average_response_time_ms": 0.0,
19+
"jokers_acquired": 1,
20+
"consumables_used": 0,
21+
"shop_visits": 13,
22+
"rerolls_used": 0,
23+
"total_money_spent": 9,
24+
"peak_money_reached": 10,
25+
"hands_played_total": 8,
26+
"discards_used_total": 0,
27+
"money_efficiency_ratio": 0.2,
28+
"resource_management_score": 0.13,
29+
"hand_type_distribution": {},
30+
"decision_consistency_score": 0.7862815887873617,
31+
"risk_assessment_score": 0.7574457126167107,
32+
"joker_synergy_score": 0.0,
33+
"failed_requests": 0,
34+
"total_reasoning_length": 5957,
35+
"average_reasoning_complexity": 397.1333333333333
36+
}

0 commit comments

Comments
 (0)