crosscodeeval_v2/README.md

CrossCodeEval v2

Dataset stats

• 6,710 v1 tasks → 863 retained; 5,847 newly archived. All 863 retain their v1 repairs; v2 changes no task instructions, environments, verifiers or reference solutions.
• 7,763 original tasks: 863 retained and 6,900 archived, including 1,053 earlier v1 exclusions.
• Images: the original 3 image definitions were not consolidated. Retained tasks reuse the 3 v1 cached images, pinned by immutable revision and checksum in the Parquet. No image copies are needed in v2.

Dataset characteristics

Cross-file code completion in C#, Java, Python and TypeScript, with related-file context and executable verifiers.

Domains

Software Engineering; Code Completion; Cross-file Repository Understanding.

Capabilities

Read related files, resolve identifiers, follow existing code patterns and complete missing statements.

Likely transfer targets: cross-file code completion and repository understanding.

Pass@16 filtering

Keep a task if either model solved it at least once. Both models completed 16 attempts on every v1 task. Recorded terminal failures count as unsuccessful attempts; unsolved tasks are not necessarily defective.

Model	Tasks evaluated	Attempts	Tasks solved at least once
Qwen3-30B Instruct	6,710	107,360	641
Qwen3-30B Thinking	6,710	107,360	627
Either model	6,710	214,720	863

Successful-attempt distribution among the 863 retained tasks. Each model column totals 863; rows with zero tasks for both models are omitted. A task with 0/16 for one model was solved by the other.

Successful attempts	Instruct tasks	Thinking tasks
0/16	222	236
1/16	222	193
2/16	70	94
3/16	57	64
4/16	40	51
5/16	38	36
6/16	33	22
7/16	33	24
8/16	20	24
9/16	22	21
10/16	17	16
11/16	21	16
12/16	18	19
13/16	17	12
14/16	13	13
15/16	7	7
16/16	13	15

The retained tasks inherit v1's rule-based repairs, build checks, reference reward 1 and no-op reward 0. Teacher runs used the same task contents without the timeout/anti-cheating sentence; v2 preserves that v1 sentence. V1 repair details · Teacher results.

Select tasks by successful-attempt count

tasks.parquet includes instruct_successes, thinking_successes and attempts_per_model (16). Counts are also available in selection.json. Excluded payloads remain in the pinned v1 source recorded in selection.json; no duplicate archive Parquet is uploaded.

import pyarrow.parquet as pq

# Exactly 1 successful attempt out of 16 for Instruct:
filters = [("instruct_successes", "=", 1)]

# Alternatively, at least 3 successes for either model:
# filters = [[("instruct_successes", ">=", 3)],
#            [("thinking_successes", ">=", 3)]]

subset = pq.read_table("tasks.parquet", filters=filters)
pq.write_table(subset, "selected.parquet")

patch.py reproduces the release from the pinned v1 Parquets and recorded counts.

Runtime

• Versions: Apptainer 1.5.4, Python 3.12.14, Harbor 0.8.1.
• Implementation: Marin Harbor 7faf878c14b6 + our runtime patches.

Traces

Trace Parquets contain 214,720 attempts, including successful and failed attempts from every evaluated task, retained or excluded. Zstandard-compressed, one row per attempt; 1 rows have no saved conversation and are marked trajectory_found=false.

Columns include task_id, model, attempt_id, reward, verifier_reward, retained_v2, trajectory_found, and the full trajectory and result JSON. reward follows the published evaluation policy; verifier_reward preserves the raw verifier score. The manifest pins the original trace source.

import json
import pyarrow.dataset as ds
traces = ds.dataset("traces", format="parquet", exclude_invalid_files=True)
failed = traces.to_table(filter=(ds.field("reward") == 0) & ds.field("trajectory_found"))
conversation = json.loads(failed["trajectory"][0].as_py())