Rockerleo commited on
Commit
12d3df4
·
verified ·
1 Parent(s): 0750799

Release har2client

Browse files
MANIFEST.json CHANGED
@@ -1311,6 +1311,18 @@
1311
  "path": "results/figures/final_cost_vs_score.png",
1312
  "sha256": "ecc0bfb6ad28d39e41a5a480eb060776a19d4945421d6e8246c7192ce325ea9f"
1313
  },
 
 
 
 
 
 
 
 
 
 
 
 
1314
  {
1315
  "path": "results/figures/heal_v3_eval_families.png",
1316
  "sha256": "03dde172c6a2f6451a04f35b47335504036badfbd8fd987db246c4396b5fd865"
 
1311
  "path": "results/figures/final_cost_vs_score.png",
1312
  "sha256": "ecc0bfb6ad28d39e41a5a480eb060776a19d4945421d6e8246c7192ce325ea9f"
1313
  },
1314
+ {
1315
+ "path": "results/figures/final_effort_matched.png",
1316
+ "sha256": "d2b3f5d9a0d983bab85c8f1f7d763fe0f64c0091cb1c2e893e6199b75409b77e"
1317
+ },
1318
+ {
1319
+ "path": "results/figures/final_all_scores.png",
1320
+ "sha256": "9d53cd1a239bcd641ebf23002da21f1e238db28f3b5542787a4f9a7266a59b13"
1321
+ },
1322
+ {
1323
+ "path": "results/figures/final_beam_by_effort.png",
1324
+ "sha256": "6f3d93ce14e70d67a7e42bb9ec5d57ac0c819e322fd4f18a4ea3f36fcdfd3346"
1325
+ },
1326
  {
1327
  "path": "results/figures/heal_v3_eval_families.png",
1328
  "sha256": "03dde172c6a2f6451a04f35b47335504036badfbd8fd987db246c4396b5fd865"
README.md CHANGED
@@ -92,6 +92,34 @@ Mean pass rate (share of hidden tests passed). Beam (high) on case 1 and case 3
92
 
93
  Beam (high) is first on three of the five (on case 1 the gap to GPT-OSS-120B is inside the interval). GPT-OSS-120B (high) is first on case 3 base and on the case 2 no-probe arm. Qwen3.5-397B-A17B is last of the three on all five.
94
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
95
  ## How to use
96
 
97
  ```python
@@ -290,6 +318,8 @@ Floor: the unchanged `client_v1.py` scores a mean of 6.5% on the 64 tasks (about
290
 
291
  The same five benchmarks and the same hidden tests for all models. Case 1 is the 50 Tier B tasks. Case 2 is the same 50 tasks, compared on the placebo arm (no probes) and the P5 arm (up to 5 probes); both arms have `submit_client`, and only P5 has `http_request`. Case 3 is the 64-task evaluation set, base and hint arms. The table shows the mean pass rate and, in brackets, the difference from Beam (high) in points with the 95% percentile bootstrap interval over tasks (paired by task, 10,000 resamples). Source: `results/analysis/model_comparison/report.md` and `summary.json`; the per-task CSVs are in `results/model_comparison/` (and `results/case2_probe/main_full/results.csv` for Beam's case 2).
292
 
 
 
293
  | Setting | Case 1 Tier B (n=50) | Case 3 base (n=64) | Case 3 + hint (n=64) | Case 2 placebo (n=50) | Case 2 P5 (n=50) |
294
  |---|---|---|---|---|---|
295
  | Beam (low) | 57.1% (-17.3 [-26.6, -9.0]) | 31.6% (-14.4 [-21.0, -8.3]) | 71.6% (-16.1 [-24.8, -7.8]) | 53.1% (+3.4 [-11.1, +17.7]) | 64.9% (-30.9 [-44.9, -17.1]) |
 
92
 
93
  Beam (high) is first on three of the five (on case 1 the gap to GPT-OSS-120B is inside the interval). GPT-OSS-120B (high) is first on case 3 base and on the case 2 no-probe arm. Qwen3.5-397B-A17B is last of the three on all five.
94
 
95
+ **All settings: low, medium and high.** Beam and GPT-OSS-120B ran at low, medium and high reasoning effort. Qwen3.5-397B-A17B has one setting (thinking on); it is not low, medium or high. Same run counts as above (Beam high on cases 1 and 3 = mean of 3 runs; every other cell = 1 run). Mean output tokens per task are in the second table; the effort names are API settings, not equal compute.
96
+
97
+ | Setting | Case 1, Tier B (50) | Case 2, no probes (50) | Case 2, 5 probes (50) | Case 3, base (64) | Case 3, with hint (64) |
98
+ |---|---|---|---|---|---|
99
+ | Beam (low) | 57.1% | 53.1% | 64.9% | 31.6% | 71.6% |
100
+ | Beam (medium) | 63.7% | 46.3% | 87.7% | 37.0% | 76.9% |
101
+ | Beam (high) | 74.5% | 49.7% | 95.7% | 46.0% | 87.7% |
102
+ | GPT-OSS-120B (low) | 51.1% | 50.0% | 62.6% | 32.7% | 47.7% |
103
+ | GPT-OSS-120B (medium) | 53.7% | 56.6% | 57.4% | 34.4% | 50.4% |
104
+ | GPT-OSS-120B (high) | 69.7% | 69.4% | 82.3% | 58.5% | 57.8% |
105
+ | Qwen3.5-397B-A17B (thinking on) | 36.0% | 40.0% | 48.0% | 29.6% | 52.3% |
106
+
107
+ | Setting | Case 1, Tier B (50) | Case 2, no probes (50) | Case 2, 5 probes (50) | Case 3, base (64) | Case 3, with hint (64) |
108
+ |---|---|---|---|---|---|
109
+ | Beam (low) | 1,154 | 1,046 | 1,046 | 1,758 | 1,800 |
110
+ | Beam (medium) | 2,259 | 2,283 | 2,110 | 2,922 | 2,846 |
111
+ | Beam (high) | 4,818 | 5,634 | 4,968 | 7,746 | 6,824 |
112
+ | GPT-OSS-120B (low) | 2,133 | 2,201 | 2,297 | 2,497 | 2,508 |
113
+ | GPT-OSS-120B (medium) | 2,183 | 2,120 | 2,229 | 2,445 | 2,490 |
114
+ | GPT-OSS-120B (high) | 5,854 | 5,402 | 5,903 | 10,734 | 10,163 |
115
+ | Qwen3.5-397B-A17B (thinking on) | 3,159 | 4,065 | 3,727 | 5,050 | 5,309 |
116
+
117
+ ![Pass rate at low, medium and high effort: Beam against GPT-OSS-120B, with Qwen3.5 as a reference line](results/figures/final_effort_matched.png)
118
+
119
+ ![Every score: 7 settings on 5 benchmarks with 95% intervals](results/figures/final_all_scores.png)
120
+
121
+ ![Beam at low, medium and high effort, with and without the missing piece](results/figures/final_beam_by_effort.png)
122
+
123
  ## How to use
124
 
125
  ```python
 
318
 
319
  The same five benchmarks and the same hidden tests for all models. Case 1 is the 50 Tier B tasks. Case 2 is the same 50 tasks, compared on the placebo arm (no probes) and the P5 arm (up to 5 probes); both arms have `submit_client`, and only P5 has `http_request`. Case 3 is the 64-task evaluation set, base and hint arms. The table shows the mean pass rate and, in brackets, the difference from Beam (high) in points with the 95% percentile bootstrap interval over tasks (paired by task, 10,000 resamples). Source: `results/analysis/model_comparison/report.md` and `summary.json`; the per-task CSVs are in `results/model_comparison/` (and `results/case2_probe/main_full/results.csv` for Beam's case 2).
320
 
321
+ The figures `results/figures/final_effort_matched.png`, `final_all_scores.png` and `final_beam_by_effort.png` (above) show every cell of this table: low, medium and high for Beam and GPT-OSS-120B, and the single thinking-on setting for Qwen3.5-397B-A17B.
322
+
323
  | Setting | Case 1 Tier B (n=50) | Case 3 base (n=64) | Case 3 + hint (n=64) | Case 2 placebo (n=50) | Case 2 P5 (n=50) |
324
  |---|---|---|---|---|---|
325
  | Beam (low) | 57.1% (-17.3 [-26.6, -9.0]) | 31.6% (-14.4 [-21.0, -8.3]) | 71.6% (-16.1 [-24.8, -7.8]) | 53.1% (+3.4 [-11.1, +17.7]) | 64.9% (-30.9 [-44.9, -17.1]) |
results/figures/final_all_scores.png ADDED

Git LFS Details

  • SHA256: 9d53cd1a239bcd641ebf23002da21f1e238db28f3b5542787a4f9a7266a59b13
  • Pointer size: 131 Bytes
  • Size of remote file: 254 kB
results/figures/final_beam_by_effort.png ADDED

Git LFS Details

  • SHA256: 6f3d93ce14e70d67a7e42bb9ec5d57ac0c819e322fd4f18a4ea3f36fcdfd3346
  • Pointer size: 131 Bytes
  • Size of remote file: 164 kB
results/figures/final_effort_matched.png ADDED

Git LFS Details

  • SHA256: d2b3f5d9a0d983bab85c8f1f7d763fe0f64c0091cb1c2e893e6199b75409b77e
  • Pointer size: 131 Bytes
  • Size of remote file: 234 kB