FWeindel commited on
Commit
7a0df9a
·
verified ·
1 Parent(s): 136122b

Endless and BugsInPy: apply resolver path fixes (validated in stage 4) and update cards

Browse files
bugsinpy_v1/README.md CHANGED
@@ -2,8 +2,11 @@
2
 
3
  ## Dataset stats
4
 
5
- - **Tasks:** 501 native bugs = **489 kept (all 489 converted/patched)** + **11 archived** + **1 not converted** (Keras 12 has identical buggy/fixed commits). Includes 462 kept instruction rewrites.
6
- - **Images:** native source supplied none; conversion introduces **11 shared image definitions**. **11 verified cached Apptainer bundles** cover all 489 kept tasks, pinned by checksum and immutable image commit in `tasks.parquet`.
 
 
 
7
 
8
  ## Dataset characteristics
9
 
@@ -27,10 +30,10 @@ Real Python repositories at buggy commits, upstream regression tests and referen
27
 
28
  Counts overlap. Build, reference and no-op checks passed before instruction review; rule-based checks reran afterwards.
29
 
30
- - **Build and setup:** **489 converted** to shared Apptainer-compatible environments with pinned dependencies; **247** include compiler/system libraries. Shared Marin/Harbor bridge fixes addressed container startup, uploads and certificate mounts. **1 archived** (`environment_task_mismatch`): YouTube-DL 40 requires older-Python behavior absent from its recorded environment.
31
- - **No-op baseline (reward 0):** **489 passed**. **6 verifier repairs:** Ansible CPU regression (1), missing-function assertions moved inside tests (4), offline YouTube-DL chapter checks (1). **3 archived** (`buggy-version-passes-verifier`): Keras 15 and Scrapy 7/11 passed without a fix.
32
- - **Reference solution (reward 1):** **489 passed**. Clean verifier dependency installs for **489**; source-timestamp fixes for **66**, plus **169 pandas** native rebuilds. **1 archived** (`flagged_by_instruction_loop`): HTTPie 4 depends on live httpbin.org availability.
33
- - **Rule-based checks:** 489 retain the timeout/anti-cheating suffix required by `check-instruction-suffix.sh` (427 carried the historical `anti-cheat-instruction` change label). **0 static archives**; 7 recorded `warning:relative-instruction-path` findings remained warnings. The contract excluded 13 checks. Separate instruction review preserved **462 rewrites** and archived **6** for tests requiring specific implementations (5) or testing a different issue (1), recorded as `flagged_by_instruction_loop`, not static-check failures.
34
 
35
  All 11 archives came from conversion or instruction review; the bullets group them by failure cause.
36
 
 
2
 
3
  ## Dataset stats
4
 
5
+ - **Original tasks:** 501
6
+ - **Kept:** 489
7
+ - **Changed:** 489
8
+ - **Archived:** 11 (plus 1 not converted: Keras 12 has identical buggy and fixed commits)
9
+ - **Images:** none supplied by the native source; 11 shared image definitions now, as 11 verified cached Apptainer bundles covering all 489 kept tasks, pinned by checksum and immutable image commit in `tasks.parquet`.
10
 
11
  ## Dataset characteristics
12
 
 
30
 
31
  Counts overlap. Build, reference and no-op checks passed before instruction review; rule-based checks reran afterwards.
32
 
33
+ - **Build and setup: 489 changed, 1 archived.** Converted to shared Apptainer-compatible environments with pinned dependencies; 247 include compiler/system libraries. Shared Marin/Harbor bridge fixes addressed container startup, uploads and certificate mounts. Archived: YouTube-DL 40 requires older-Python behavior absent from its recorded environment.
34
+ - **No-op baseline (reward 0): 6 changed, 3 archived.** Verifier repairs: Ansible CPU regression (1), missing-function assertions moved inside tests (4), offline YouTube-DL chapter checks (1). Archived: Keras 15 and Scrapy 7/11 passed without a fix.
35
+ - **Reference solution (reward 1): 489 changed, 1 archived.** Clean verifier dependency installs for 489; source-timestamp fixes for 66, plus 169 pandas native rebuilds. Archived: HTTPie 4 depends on live httpbin.org availability.
36
+ - **Rule-based checks: 462 changed, 6 archived.** A separate instruction review kept 462 rewrites. Archived: tests requiring specific implementations (5) or testing a different issue (1). 7 kept tasks were flagged for relative paths; the path resolver made 1 path absolute (bugsinpy-original-luigi-9). The other 6 keep advisory warnings; all were inspected and are intended (example names, command arguments, URL and traceback fragments, one return value). Every instruction also ends with "You have N seconds to complete this task. Do not cheat by using online solutions or hints specific to this task." (N is the agent timeout); this sentence is not counted as a change.
37
 
38
  All 11 archives came from conversion or instruction review; the bullets group them by failure cause.
39
 
bugsinpy_v1/tasks.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:0c9b90adbe2e2d5ac8f56165ebfaa5e3dbde370f51b453884be5b3c0b2944c54
3
- size 4343661074
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ca7d89da898adc14f5ce547b49552197d726010e89282ffe30024bda949a636c
3
+ size 4343201142
endless_v1/README.md CHANGED
@@ -2,7 +2,11 @@
2
 
3
  ## Dataset stats
4
 
5
- **2,492 original tasks; 2,144 kept, all 2,144 patched; 348 archived.** Original: **2,492 unique build contexts**. Current: **1,834 cached Apptainer environments** covering all kept tasks; 310 fewer environments than kept tasks.
 
 
 
 
6
 
7
  ## Dataset characteristics
8
 
@@ -21,10 +25,10 @@ Intended benchmark transfer: **Terminal-Bench 2.0 and OpenThoughts-TBLite 2.0**,
21
 
22
  ## Per-stage breakdown
23
 
24
- - **Build and setup: 2,144 kept tasks passed.** All **2,492 original contexts** built successfully. Kept repairs: **1,721** build compatibility, **419** shared dependencies/setup, **4** startup compatibility; all **2,144** use offline verification. **9** timestamp repairs overlap. **271 archived** for reviewed identity, ownership or permission mismatches. Conversion review covered **435 tasks**, without a Docker reference run.
25
- - **No-op baseline (reward 0): 2,144 kept tasks passed.** **23 archived:** 22 false successes and 1 verifier repeatedly writing reward 1 without tests.
26
- - **Reference solution (reward 1): 2,144 kept tasks passed.** Repairs include pinned pip installs for **8 tasks**, CPU handling for **4**, and subprocess PATH for **1**; **53 archived** after build exclusions.
27
- - **Rule-based checks:** restored timeout/anti-cheating suffixes for **2,144 kept tasks**; **1 archived for `reference-solution-integrity`**, a fake YAML module accepted by version-only grading. **3 kept truncated prompts** had already been restored from source descriptions.
28
 
29
  Repair counts overlap; each archive is counted once. Runtime results reuse historical full runs and repair pilots. One HTTP task retains an unresolved possible shared-port collision.
30
 
 
2
 
3
  ## Dataset stats
4
 
5
+ - **Original tasks:** 2,492
6
+ - **Kept:** 2,144
7
+ - **Changed:** 2,144
8
+ - **Archived:** 348
9
+ - **Images:** 2,492 original (one build context per task); 1,834 cached Apptainer environments now, covering all kept tasks.
10
 
11
  ## Dataset characteristics
12
 
 
25
 
26
  ## Per-stage breakdown
27
 
28
+ - **Build and setup: 2,144 changed, 271 archived.** All 2,492 original contexts built. Repairs: 1,721 build compatibility, 419 shared dependencies/setup, 4 startup compatibility, 9 timestamp repairs; all 2,144 use offline verification. Conversion review covered 435 tasks, without a Docker reference run. Archived: reviewed identity, ownership or permission mismatches.
29
+ - **No-op baseline (reward 0): 0 changed, 23 archived.** Archived: 22 false successes and 1 verifier repeatedly writing reward 1 without tests.
30
+ - **Reference solution (reward 1): 12 changed, 53 archived.** Repairs: pinned pip installs (8), CPU handling (4), subprocess PATH (1). Archived after build exclusions.
31
+ - **Rule-based checks: 74 changed, 1 archived.** 3 truncated instructions restored from source descriptions. Archived: a fake YAML module accepted by version-only grading. 136 kept tasks were flagged for relative paths in the instruction; the path resolver made 161 paths absolute in 71 of them. A random sample of 50 showed only intended relative mentions: values the output must contain (21), paths relative to a named directory (15), required relative symlink targets (10) and paths inside archives (4). Every instruction also ends with "You have N seconds to complete this task. Do not cheat by using online solutions or hints specific to this task." (N is the agent timeout); this sentence is not counted as a change.
32
 
33
  Repair counts overlap; each archive is counted once. Runtime results reuse historical full runs and repair pilots. One HTTP task retains an unresolved possible shared-port collision.
34
 
endless_v1/tasks.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:8796986aa27e93b2b8cb79b37a20cb146008e3d466f371a5c7fa8998bd433c3c
3
- size 31822240
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5081aca34db5733eed66fd4774a4587edee4f092e561d9266a874d2f7ae2475c
3
+ size 76024142