crosscodeeval_v2/README.md

CrossCodeEval v2

Dataset stats

• v1 tasks evaluated: 6,710
• Kept: 863 (solved at least once by either model)
• Excluded: 5,847 (no recorded success)
• Attempts: pass@16 with Qwen3-30B Instruct and Qwen3-30B Thinking (214,720 in total)

Evaluation settings

Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget.

Model	Context tokens (configured)	Intended response tokens per call	Output-limit evidence	Thinking	Sampling
openai/qwen3-30b-instruct-2507	32768	8192	No explicit request cap in sampled saved configs; actual fallback cap unknown	off	{"min_p": 0, "temperature": 0.7, "top_k": 20, "top_p": 0.8}
openai/qwen3-30b-thinking	32768	8192	No explicit request cap in sampled saved configs; actual fallback cap unknown	on	{"min_p": 0, "temperature": 0.6, "top_k": 20, "top_p": 0.95}

Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults.

Historical output-limit bug: the intended per-response limit was stored as top-level max_tokens, but the affected Terminus path did not forward it as llm_call_kwargs.max_tokens. Thus these sampled runs did not request the intended explicit cap. Generation instead depended on the backend's fallback allowance, available context, stopping conditions and agent timeout. The exact fallback limit is not established by the saved configurations; it must not be reported as 8,192 or 16,384 enforced tokens, or as unlimited.

The table reports settings found in 6 sampled saved trial configurations, not an exhaustive audit of every attempt or direct proof of server settings. Historical runs and retries may differ. Sampling values are recorded explicitly rather than inferred from today's model-card defaults. Agent and verifier wall-clock budgets remain task-specific; sampled resolved limits are in settings and evidence. Context is not the sum of tokens across all turns. The current forwarding fix does not retroactively change these pass@k results.

Pass@16 filtering

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 attempts among kept tasks

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

Select tasks by successful-attempt count

You can select tasks by how often they were solved, for example all tasks Instruct solved exactly once:

import pyarrow.parquet as pq

subset = pq.read_table("tasks.parquet", filters=[("instruct_successes", "=", 1)])

Traces

traces/ holds all 214,720 recorded attempts, successful and failed, from every evaluated task, one row per attempt (Zstandard-compressed Parquet). Columns include task_id, model, attempt_id, reward, verifier_reward, retained_v2 and the full trajectory and result JSON. 1 attempts have no saved conversation (trajectory_found=false).

import json
import pyarrow.dataset as ds

traces = ds.dataset("traces", format="parquet")
failed = traces.to_table(filter=ds.field("reward") == 0)
conversation = json.loads(failed["trajectory"][0].as_py())

Trajectory statistics (plots, command distribution) are in trace_stats.

Not yet checked with the repository-leak checks added on 2026-10-17 (leftover git history and file timestamps that could reveal the solution).

Runtime

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

patch.py rebuilds tasks.parquet from the current v1 tasks and the recorded results, both pinned to revision a43f5635.