FWeindel's picture
Standardize v2 and v3 READMEs and pass@k plots
7bd74c0 verified
|
Raw History Blame Contribute Delete
4.99 kB

NL2Bash v3

Dataset stats

  • Tasks: 487 had no successful attempt in v2; all 487 evaluated for v3; 104 newly kept; 383 still without success.
  • Teacher evaluation: Qwen/Qwen3.8-27B (thinking off), 2 attempts per task (pass@2).
  • Coverage: 487 tasks evaluated; 974 evaluation attempts.
  • Saved conversations: 974 of 974 published attempt records, including successful and unsuccessful attempts.

We keep tasks with at least one verifier-successful attempt from the new model. tasks.parquet contains only newly kept tasks; tasks already kept in v2 remain separate.

Evaluation settings

Model Agent Thinking Attempts/task Temperature Top-p Top-k Context tokens Intended output tokens/call Other sampling
Qwen/Qwen3.8-27B Terminus-2 Off 2 0.7 0.8 20 32,768 8,192 presence_penalty=1.5; repetition_penalty=1

Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total.

Time limits: agent 600 seconds; verifier 600 seconds (all 974 saved configurations; limits vary by task where multiple values are shown).

Output-limit evidence: All 974 saved configurations contain an explicit 8,192-token request cap. HTTP payloads and server enforcement were not audited.

Network isolation was requested, but the run fell back to host networking when network namespace creation failed; agents could have reached the network.

Results

The tables and plot cover all evaluated tasks for each model, before selecting tasks for publication. Only the new v3 evaluation is included.

Observed success is a solvability filter, not proof of verifier correctness. Inspected zero-score examples included output-format mismatches and faulty expected output.

Successful attempts Qwen/Qwen3.8-27B (487 tasks)
0/2 383
1/2 51
2/2 53
Model Evaluated tasks Pass@1 Pass@2
Qwen/Qwen3.8-27B 487 16.12% 21.36%

Pass@k uses the sampling-without-replacement estimator from the recorded per-task success counts. At the full attempt budget, it is the fraction of evaluated tasks with at least one observed success. Recorded terminal errors follow the saved scoring policy; absent attempts are not fabricated and raw missing verifier rewards remain null.

Pass rates across all evaluated tasks

Files and usage

tasks.parquet contains 104 newly kept tasks only. Tasks already kept in v2 are not duplicated.

traces/ contains attempt records across the evaluated cohort, including unsuccessful attempts and tasks not kept. Rows retain task/model/attempt identifiers, raw results and conversations when available.

Success-count columns: coder_successes, qwen38_successes. Counts are model-specific. The settings above describe this release's evaluation; historical columns retain their earlier budgets.

archive.parquet preserves excluded task payloads and reasons, including earlier exclusions where applicable.

from huggingface_hub import hf_hub_download
import pyarrow.parquet as pq

path = hf_hub_download(
    "FWeindel/validated-tasks",
    "nl2bash_v3/tasks.parquet",
    repo_type="dataset",
)
# Select kept tasks with exactly one observed success.
tasks = pq.read_table(path, filters=[("qwen38_successes", "=", 1)])

Combine the 938 tasks in nl2bash_v2/tasks.parquet with these 104 newly kept tasks for 1,042 unique tasks. Allow for different columns and retain each source file's image manifest when staging images.

import pyarrow as pa

def load(version):
    path = hf_hub_download(
        "FWeindel/validated-tasks",
        f"nl2bash_{version}/tasks.parquet",
        repo_type="dataset",
    )
    return pq.read_table(path)

v2, v3 = load("v2"), load("v3")
assert set(v2["path"].to_pylist()).isdisjoint(v3["path"].to_pylist())
combined = pa.concat_tables([v2, v3], promote_options="permissive")

Repository-history and timestamp leak checks remain incomplete for these tasks. Teacher success does not establish completion of all current datasource-validation checks.

Trace statistics

Of 974 recorded attempts, 974 have saved conversations and 0 do not. 1 attempt has no raw verifier reward. Missing measurements remain missing.

Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade.

The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. Historical attempts remain in v2.

trace_stats/README.md contains detailed turns, token usage, context, runtime, outcome and command distributions. Token totals sum calls across an attempt and are not the context window.