Drop pandas__49324-49704 (ungradable PR-1 install), 14 tasks

README.md

@@ -13,53 +13,55 @@ configs:
 
 # foresight-swe-mini
 
-15 of the 212 tasks in [Foresight-SWE](https://huggingface.co/datasets/shaswat123/Foresight-SWE), chosen to match the full benchmark's mix of languages, difficulty and defect types while keeping runs short.
+14 of the 212 tasks in [Foresight-SWE](https://huggingface.co/datasets/shaswat123/Foresight-SWE), chosen to match the full benchmark's mix of languages, difficulty and defect types while keeping runs short.
 
 Foresight-SWE asks whether a coding agent avoids a bug nobody told it about. Each task pairs a real pull request that introduced a defect (PR-1) with the later one that fixed it (PR-2). The agent gets only PR-1's task. Resolve rate counts PR-1's tests passing. Foresight resolve rate (FRR) also requires PR-2's held-out tests to pass. See the [paper](https://arxiv.org/abs/2611.01181) and [code](https://github.com/shaswatpatel123/Foresight-SWE).
 
-**Read this first.** Unlike the other mini benchmarks in this collection, this subset has not been checked against real per-task results, because upstream publishes only totals and per-model breakdowns. The accuracy figures below are simulated. The tasks have not been run through the upstream harness here yet, so a task may be swapped if it turns out to be broken or very slow. Any change will be noted on this card.
+**Read this first.** Unlike the other mini benchmarks in this collection, this subset has not been checked against real per-task results, because upstream publishes only totals and per-model breakdowns. The accuracy figures below are simulated.
+
+**Changes.** The subset started with 15 tasks. Running upstream's gold patches through the upstream harness passed PR-1's tests on 14. The 15th, `pandas__49324-49704`, cannot be graded: its PR-1 install recipe installs numpy 2 before building pandas, the build fails, and no PR-1 test runs even for the gold patch. It was removed, and every number on this card is for the remaining 14. The gold patches also pass PR-2's held-out tests on 1 of the 14, so a patch identical to upstream's PR-1 would score 1 of 14 on FRR.
 
 ## What it can tell you
 
-One task is worth 6.7 points, and at leaderboard rates most models would pass 1 to 4 of the 15. Use it to spot large differences, or to compare two setups on the same tasks. Do not use it to rank close models or to estimate a full-benchmark score.
+One task is worth 7.1 points, and at leaderboard rates most models would pass 1 to 4 of the 14. Use it to spot large differences, or to compare two setups on the same tasks. Do not use it to rank close models or to estimate a full-benchmark score.
 
 Simulated against the 11 models on the leaderboard:
 
 | tasks | typical FRR error (points) | pairs 5+ points apart in the right order | pairs 10+ points apart in the right order |
 |---:|---:|---:|---:|
-| **foresight-swe-mini (15)** | **6.8** | **74%** | **83%** |
-| random 10 | 8.8 | 64% | 72% |
-| random 15 | 6.9 | 73% | 82% |
-| random 20 | 6.0 | 78% | 87% |
-| random 30 | 4.7 | 84% | 93% |
-| random 60 | 3.0 | 94% | 99% |
+| **foresight-swe-mini (14)** | **6.9** | **73%** | **82%** |
+| random 10 | 8.6 | 64% | 73% |
+| random 14 | 7.0 | 72% | 82% |
+| random 20 | 5.9 | 79% | 88% |
+| random 30 | 4.7 | 85% | 94% |
+| random 60 | 3.1 | 95% | 99% |
 | random 100 | 2.0 | 98% | 100% |
 
-The 15 do no better than a random 15 on these numbers. The published task labels explain only 4 to 13% of each model's pass or fail outcomes, so matching on them cannot do much for accuracy. What matching adds is a guaranteed mix and no very slow tasks. A subset chosen from real per-task results might do better; that needs data upstream has not released.
+The 14 do no better than a random 14 on these numbers. The published task labels explain only 4 to 13% of each model's pass or fail outcomes, so matching on them cannot do much for accuracy. What matching adds is a guaranteed mix and no very slow tasks. A subset chosen from real per-task results might do better; that needs data upstream has not released.
 
 The simulation fits each model to Tables D.1 and D.2 of the paper, then adds a difficulty shared across models and noise per model. Those two additions are assumptions. Numbers are the middle of three runs with different shared-difficulty settings, which changed little.
 
 ## The mix
 
-| | foresight-swe-mini | exact share of 15 |
+| | foresight-swe-mini | exact share of 14 |
 |---|---:|---:|
-| Python / JavaScript / TypeScript / Go | 8 / 3 / 2 / 2 | 8.1 / 3.4 / 2.0 / 1.5 |
-| PR-1 easy / medium / hard | 3 / 6 / 6 | 3.3 / 5.4 / 6.2 |
-| PR-2 easy / medium / hard | 9 / 4 / 2 | 8.8 / 3.7 / 2.5 |
-| Major / minor | 9 / 6 | 9.1 / 5.9 |
-| Silent / loud | 9 / 6 | 9.1 / 5.9 |
-| Bug / feature / mixed | 8 / 6 / 1 | 7.9 / 6.2 / 0.8 |
-| Security | 1 | 1.3 |
+| Python / JavaScript / TypeScript / Go | 7 / 3 / 2 / 2 | 7.6 / 3.2 / 1.8 / 1.4 |
+| PR-1 easy / medium / hard | 2 / 6 / 6 | 3.1 / 5.1 / 5.8 |
+| PR-2 easy / medium / hard | 9 / 3 / 2 | 8.3 / 3.4 / 2.3 |
+| Major / minor | 9 / 5 | 8.5 / 5.5 |
+| Silent / loud | 9 / 5 | 8.5 / 5.5 |
+| Bug / feature / mixed | 7 / 6 / 1 | 7.4 / 5.8 / 0.8 |
+| Security | 1 | 1.2 |
 
-The 15 cover 10 of the 35 repositories, with at most 3 tasks from any one. Predicted FRR on the subset is within 0.6 points of the published score for all 11 models (`data/leaderboard.csv`). That only checks the mix, not real results.
+The 14 cover 10 of the 35 repositories, with at most 3 tasks from any one. Predicted FRR on the subset is within 1.6 points of the published score for all 11 models (`data/leaderboard.csv`). That only checks the mix, not real results.
 
-## How the 15 were chosen
+## How the tasks were chosen
 
-Simulated annealing over the 158 tasks with at most 2,000 regression tests, 100 task tests, 20 held-out tests, a 60 KB patch and 3.5 GB of images. It minimises the difference from the full benchmark in the mix above, in each model's predicted scores and in task size, with a cap per repository. `python build.py verified.jsonl` reproduces the selection and every number here. Exact task IDs are in `data/tasks.jsonl`.
+Simulated annealing over the 158 tasks with at most 2,000 regression tests, 100 task tests, 20 held-out tests, a 60 KB patch and 3.5 GB of images. It minimises the difference from the full benchmark in the mix above, in each model's predicted scores and in task size, with a cap per repository. The search picked 15; the broken task was then dropped. `python build.py verified.jsonl` reproduces the selection, the removal and every number here. Exact task IDs are in `data/tasks.jsonl`.
 
 ## Running it
 
-The tasks live upstream; this dataset only tells you which 15 to run. Build a task file:
+The tasks live upstream; this dataset only tells you which 14 to run. Build a task file:
 
 ```python
 from datasets import load_dataset
@@ -68,9 +70,9 @@ tasks = load_dataset("shaswat123/Foresight-SWE", split="test")
 tasks.filter(lambda r: r["instance_id"] in ids).to_json("foresight_mini.jsonl")
 ```
 
-Then set `DATASET` to that file in one of the [upstream eval scripts](https://github.com/shaswatpatel123/Foresight-SWE/tree/main/eval). Each task needs one amd64 Docker image, about 11 GB in total. To compare with the leaderboard, keep the paper's budgets: 250 to 350 steps and $3 per task in mini-swe-agent.
+Then set `DATASET` to that file in one of the [upstream eval scripts](https://github.com/shaswatpatel123/Foresight-SWE/tree/main/eval). Each task needs one amd64 Docker image, about 10 GB in total. To compare with the leaderboard, keep the paper's budgets: 250 to 350 steps and $3 per task in mini-swe-agent.
 
-Report resolve rate and FRR, and label results "foresight-swe-mini". Scores on the 15 are not Foresight-SWE scores.
+Report resolve rate and FRR, and label results "foresight-swe-mini". Scores on the 14 are not Foresight-SWE scores.
 
 ## Credit
 

build.py

@@ -101,6 +101,10 @@ def search(pool, seed, iters=25000):
 # Keep runs short: leave out the tasks that are slow to grade.
 pool = np.where((p2p <= 2000) & (ant <= 20) & (f2p <= 100) & (patch_kb <= 60) & (both_images_gb <= 3.5))[0]
 S = min((search(pool, s) for s in range(8)), key=lambda x: x[1])[0]
+# Removed after selection: upstream's PR-1 install recipe resolves numpy 2 before
+# building pandas, so the build fails and no PR-1 test can run, even for the gold patch.
+BROKEN = {"pandas__49324-49704"}
+S = [i for i in S if ids[i] not in BROKEN]
 
 
 def calibrate(lin, target):
@@ -120,7 +124,7 @@ def simulate(sd, reps=500, seed=7):
             lin[:, m] += calibrate(lin[:, m], FULL_FRR[m] / 100)
         Y = (rng.random((N, 11)) < sig(lin)) * 1.
         full = Y.mean(0)
-        designs = {f"random {k}": rng.choice(N, k, replace=False) for k in (10, 15, 20, 30, 60, 100)}
+        designs = {f"random {k}": rng.choice(N, k, replace=False) for k in (10, len(S), 20, 30, 60, 100)}
         designs["foresight-swe-mini"] = np.array(S)
         for name, D in designs.items():
             e = Y[D].mean(0)
@@ -157,7 +161,7 @@ metrics = dict(
     variance_explained_frr=[round(float(x), 3) for x in pF.var(0) / (pF.mean(0) * (1 - pF.mean(0)))],
     spearman_resolve_frr_all=round(float(spearmanr(FULL_RESOLVE, FULL_FRR)[0]), 2),
     mix={k: dict(mini=dict(C.Counter(str(x) for x in v[S])),
-                 proportional={str(a): round(b / N * K, 1) for a, b in C.Counter(v.tolist()).items()})
+                 proportional={str(a): round(b / N * len(S), 1) for a, b in C.Counter(v.tolist()).items()})
          for k, v in mix.items()},
     simulation={str(sd): simulate(sd) for sd in (.5, 1, 1.5)})
 json.dump(metrics, open("data/metrics.json", "w"), indent=1)

data/leaderboard.csv

@@ -1,12 +1,12 @@
 model,resolve_all_212,frr_all_212,predicted_frr_on_mini
-Qwen3.8 Flash Next,91.98,20.75,21.2
-GLM-5.3 Flash,80.66,11.32,11.3
-DeepSeek-V4 Flash,58.49,18.87,19.4
-GPT-5.6 Sol,84.43,18.4,19.0
-GPT-5.6 Terra,74.53,15.09,15.2
-GPT-5.6 Luna,74.06,16.51,16.9
-Claude Opus 5,77.83,24.53,24.8
-Claude Sonnet 5,59.91,16.98,17.2
-Qwen3.5-397B-A17B,21.7,7.08,6.7
-Qwen3.8-27B,25.47,7.55,7.4
-Qwen3.5-27B,18.4,6.13,6.0
+Qwen3.8 Flash Next,91.98,20.75,21.4
+GLM-5.3 Flash,80.66,11.32,10.2
+DeepSeek-V4 Flash,58.49,18.87,18.8
+GPT-5.6 Sol,84.43,18.4,19.3
+GPT-5.6 Terra,74.53,15.09,15.1
+GPT-5.6 Luna,74.06,16.51,16.5
+Claude Opus 5,77.83,24.53,22.9
+Claude Sonnet 5,59.91,16.98,16.5
+Qwen3.5-397B-A17B,21.7,7.08,6.4
+Qwen3.8-27B,25.47,7.55,6.8
+Qwen3.5-27B,18.4,6.13,5.8

data/metrics.json

@@ -1,10 +1,10 @@
 {
  "n_pool": 158,
- "image_gb": 11.0,
- "max_pass_to_pass": 1872,
+ "image_gb": 10.1,
+ "max_pass_to_pass": 752,
  "repos": 10,
- "predicted_frr_max_error": 0.6,
- "predicted_resolve_max_error": 0.9,
+ "predicted_frr_max_error": 1.6,
+ "predicted_resolve_max_error": 2.9,
  "variance_explained_frr": [
   0.132,
   0.046,
@@ -22,82 +22,82 @@
  "mix": {
   "language": {
    "mini": {
-    "python": 8,
+    "python": 7,
     "go": 2,
     "javascript": 3,
     "typescript": 2
    },
    "proportional": {
-    "python": 8.1,
-    "typescript": 2.0,
-    "javascript": 3.4,
-    "go": 1.5
+    "python": 7.6,
+    "typescript": 1.8,
+    "javascript": 3.2,
+    "go": 1.4
    }
   },
   "pr1_difficulty": {
    "mini": {
     "medium": 6,
     "hard": 6,
-    "easy": 3
+    "easy": 2
    },
    "proportional": {
-    "hard": 6.2,
-    "easy": 3.3,
-    "medium": 5.4
+    "hard": 5.8,
+    "easy": 3.1,
+    "medium": 5.1
    }
   },
   "pr2_difficulty": {
    "mini": {
-    "medium": 4,
+    "medium": 3,
     "easy": 9,
     "hard": 2
    },
    "proportional": {
-    "medium": 3.7,
-    "easy": 8.8,
-    "hard": 2.5
+    "medium": 3.4,
+    "easy": 8.3,
+    "hard": 2.3
    }
   },
   "severity": {
    "mini": {
     "major": 9,
-    "minor": 6
+    "minor": 5
    },
    "proportional": {
-    "major": 9.1,
-    "minor": 5.9
+    "major": 8.5,
+    "minor": 5.5
    }
   },
   "detectability": {
    "mini": {
     "silent": 9,
-    "loud": 6
+    "loud": 5
    },
    "proportional": {
-    "silent": 9.1,
-    "loud": 5.9
+    "silent": 8.5,
+    "loud": 5.5
    }
   },
   "task_type": {
    "mini": {
     "feature": 6,
-    "bug": 8,
+    "bug": 7,
     "mixed": 1
    },
    "proportional": {
-    "feature": 6.2,
-    "bug": 7.9,
+    "feature": 5.8,
+    "bug": 7.4,
     "mixed": 0.8
    }
   },
   "security": {
    "mini": {
     "True": 1,
-    "False": 14
+    "False": 13
    },
    "proportional": {
-    "False": 13.7,
-    "True": 1.3
+    "False": 12.8,
+    "True": 1.2
    }
   }
  },
@@ -105,112 +105,112 @@
   "0.5": {
    "random 10": {
     "error_points": 8.7,
-    "pairs_5pt": 63.6,
-    "pairs_10pt": 72.0
+    "pairs_5pt": 62.6,
+    "pairs_10pt": 70.2
    },
-   "random 15": {
-    "error_points": 6.8,
-    "pairs_5pt": 72.4,
-    "pairs_10pt": 81.6
+   "random 14": {
+    "error_points": 7.0,
+    "pairs_5pt": 72.1,
+    "pairs_10pt": 81.8
    },
    "random 20": {
     "error_points": 5.9,
-    "pairs_5pt": 77.6,
-    "pairs_10pt": 86.7
+    "pairs_5pt": 78.2,
+    "pairs_10pt": 87.5
    },
    "random 30": {
     "error_points": 4.7,
-    "pairs_5pt": 83.8,
+    "pairs_5pt": 84.5,
     "pairs_10pt": 92.8
    },
    "random 60": {
-    "error_points": 3.0,
+    "error_points": 3.1,
     "pairs_5pt": 93.8,
-    "pairs_10pt": 98.6
+    "pairs_10pt": 98.8
    },
    "random 100": {
     "error_points": 2.0,
-    "pairs_5pt": 98.1,
+    "pairs_5pt": 98.2,
     "pairs_10pt": 99.9
    },
    "foresight-swe-mini": {
-    "error_points": 6.7,
-    "pairs_5pt": 73.8,
-    "pairs_10pt": 83.6
+    "error_points": 6.9,
+    "pairs_5pt": 72.5,
+    "pairs_10pt": 81.2
    }
   },
   "1": {
    "random 10": {
-    "error_points": 8.8,
-    "pairs_5pt": 63.8,
-    "pairs_10pt": 72.5
+    "error_points": 8.6,
+    "pairs_5pt": 63.7,
+    "pairs_10pt": 72.7
    },
-   "random 15": {
-    "error_points": 6.9,
-    "pairs_5pt": 73.0,
-    "pairs_10pt": 82.3
+   "random 14": {
+    "error_points": 7.0,
+    "pairs_5pt": 71.8,
+    "pairs_10pt": 81.6
    },
    "random 20": {
-    "error_points": 6.0,
-    "pairs_5pt": 78.1,
-    "pairs_10pt": 87.1
+    "error_points": 5.9,
+    "pairs_5pt": 78.7,
+    "pairs_10pt": 88.4
    },
    "random 30": {
     "error_points": 4.7,
-    "pairs_5pt": 84.1,
-    "pairs_10pt": 92.6
+    "pairs_5pt": 85.0,
+    "pairs_10pt": 93.7
    },
    "random 60": {
-    "error_points": 3.0,
-    "pairs_5pt": 94.3,
-    "pairs_10pt": 99.0
+    "error_points": 3.1,
+    "pairs_5pt": 94.5,
+    "pairs_10pt": 99.3
    },
    "random 100": {
     "error_points": 2.0,
-    "pairs_5pt": 98.3,
-    "pairs_10pt": 100.0
+    "pairs_5pt": 98.4,
+    "pairs_10pt": 99.9
    },
    "foresight-swe-mini": {
-    "error_points": 6.8,
-    "pairs_5pt": 74.3,
-    "pairs_10pt": 83.3
+    "error_points": 6.9,
+    "pairs_5pt": 72.9,
+    "pairs_10pt": 82.1
    }
   },
   "1.5": {
    "random 10": {
-    "error_points": 8.7,
-    "pairs_5pt": 63.6,
-    "pairs_10pt": 72.3
+    "error_points": 8.6,
+    "pairs_5pt": 64.3,
+    "pairs_10pt": 74.2
    },
-   "random 15": {
-    "error_points": 6.9,
-    "pairs_5pt": 73.8,
-    "pairs_10pt": 83.2
+   "random 14": {
+    "error_points": 7.1,
+    "pairs_5pt": 72.2,
+    "pairs_10pt": 82.7
    },
    "random 20": {
-    "error_points": 6.0,
-    "pairs_5pt": 79.1,
-    "pairs_10pt": 88.2
+    "error_points": 5.9,
+    "pairs_5pt": 79.7,
+    "pairs_10pt": 89.3
    },
    "random 30": {
     "error_points": 4.7,
-    "pairs_5pt": 85.3,
-    "pairs_10pt": 93.4
+    "pairs_5pt": 85.7,
+    "pairs_10pt": 94.4
    },
    "random 60": {
     "error_points": 3.0,
-    "pairs_5pt": 94.8,
-    "pairs_10pt": 99.3
+    "pairs_5pt": 94.9,
+    "pairs_10pt": 99.4
    },
    "random 100": {
     "error_points": 2.0,
-    "pairs_5pt": 98.6,
-    "pairs_10pt": 100.0
+    "pairs_5pt": 98.7,
+    "pairs_10pt": 99.9
    },
    "foresight-swe-mini": {
-    "error_points": 6.8,
-    "pairs_5pt": 74.0,
-    "pairs_10pt": 83.3
+    "error_points": 7.0,
+    "pairs_5pt": 72.9,
+    "pairs_10pt": 82.8
    }
   }
  }

data/tasks.jsonl

@@ -1,7 +1,6 @@
 {"instance_id": "aiohttp__7665-8078", "repo": "aio-libs/aiohttp", "language": "python", "earlier_pr": 7665, "later_pr": 8078, "pr1_difficulty": "medium", "difficulty": "medium", "severity": "major", "failure_mode": "silent", "pr1_task_type": "feature", "security": true, "docker_image": "docker.io/shaswatpatel123/foresight-swe:inst-aiohttp__7665-8078-pr1", "n_fail_to_pass": 2, "n_anticipation": 6, "n_pass_to_pass": 218, "image_gb": 0.53}
 {"instance_id": "django__21204-34570", "repo": "django/django", "language": "python", "earlier_pr": 21204, "later_pr": 34570, "pr1_difficulty": "hard", "difficulty": "easy", "severity": "minor", "failure_mode": "loud", "pr1_task_type": "bug", "security": false, "docker_image": "docker.io/shaswatpatel123/foresight-swe:inst-django__21204-34570-pr1", "n_fail_to_pass": 3, "n_anticipation": 1, "n_pass_to_pass": 312, "image_gb": 0.61}
 {"instance_id": "django__34429-35477", "repo": "django/django", "language": "python", "earlier_pr": 34429, "later_pr": 35477, "pr1_difficulty": "medium", "difficulty": "easy", "severity": "major", "failure_mode": "silent", "pr1_task_type": "feature", "security": false, "docker_image": "docker.io/shaswatpatel123/foresight-swe:inst-django__34429-35477-pr1", "n_fail_to_pass": 7, "n_anticipation": 3, "n_pass_to_pass": 164, "image_gb": 0.57}
-{"instance_id": "pandas__49324-49704", "repo": "pandas-dev/pandas", "language": "python", "earlier_pr": 49324, "later_pr": 49704, "pr1_difficulty": "easy", "difficulty": "medium", "severity": "minor", "failure_mode": "loud", "pr1_task_type": "bug", "security": false, "docker_image": "docker.io/shaswatpatel123/foresight-swe:inst-pandas__49324-49704-pr1", "n_fail_to_pass": 39, "n_anticipation": 1, "n_pass_to_pass": 1872, "image_gb": 0.9}
 {"instance_id": "pandas__62843-65342", "repo": "pandas-dev/pandas", "language": "python", "earlier_pr": 62843, "later_pr": 65342, "pr1_difficulty": "medium", "difficulty": "easy", "severity": "minor", "failure_mode": "loud", "pr1_task_type": "mixed", "security": false, "docker_image": "docker.io/shaswatpatel123/foresight-swe:inst-pandas__62843-65342-pr1", "n_fail_to_pass": 6, "n_anticipation": 2, "n_pass_to_pass": 752, "image_gb": 0.79}
 {"instance_id": "prometheus__14513-14598", "repo": "prometheus/prometheus", "language": "go", "earlier_pr": 14513, "later_pr": 14598, "pr1_difficulty": "hard", "difficulty": "hard", "severity": "minor", "failure_mode": "loud", "pr1_task_type": "bug", "security": false, "docker_image": "docker.io/shaswatpatel123/foresight-swe:inst-prometheus__14513-14598-pr1", "n_fail_to_pass": 2, "n_anticipation": 1, "n_pass_to_pass": 269, "image_gb": 1.34}
 {"instance_id": "pydantic__10347-10358", "repo": "pydantic/pydantic", "language": "python", "earlier_pr": 10347, "later_pr": 10358, "pr1_difficulty": "easy", "difficulty": "easy", "severity": "minor", "failure_mode": "loud", "pr1_task_type": "bug", "security": false, "docker_image": "docker.io/shaswatpatel123/foresight-swe:inst-pydantic__10347-10358-pr1", "n_fail_to_pass": 1, "n_anticipation": 1, "n_pass_to_pass": 44, "image_gb": 0.68}