pranay5255/agent-trajectory-study-20260930

Agent trajectory study

This corpus contains 1,451 recorded episodes: 570 OLMo, 510 parseable Qwen, and 371 forestOfAudits episodes. It retains companion views and rejected OLMo responses, with provenance and compressed recordings. Known copies share an episode identity; separate attempts with identical bytes remain separate episodes.

Qwen recordings concern EVMBench detect tasks in the Harbor/SkyRL pipeline. Their collection stages are diagnostics, stress tests, and generation-only admission gates. The presence of rewards and admission checks does not establish that RL updates occurred. OLMo checkpoint labels describe released model stages. Comparisons across stages, harnesses, and tasks are observational.

Run the Colab notebook

Download trajectory_eda_colab.ipynb. In Google Colab, choose File > Upload notebook, upload the file, and use the standard CPU runtime. Run the cells in order. No Hugging Face login, GPU, or paid model calls are needed. Colab's own account requirements still apply.

Compact episode and metadata tables load first. The interactive inspector downloads one compressed episode bundle at a time. The notebook provides coverage, size and usage, outcome denominators, command indicators, collection-specific evidence, task-bootstrap intervals, batched selected-token diagnostics, event search, and downloadable CSVs and figures. It records the resolved Hub revision and package versions.

Download through the CLI:

hf download pranay5255/agent-trajectory-study-20260930 trajectory_eda_colab.ipynb --repo-type dataset --local-dir .

For a local copy, set CORPUS_LOCAL_ROOT to the repository directory. To pin a published snapshot, set CORPUS_REVISION to its commit hash before running the notebook.

Tables and row units

Configuration	Row unit	Meaning
olmo, qwen, forest	Episode	One audited recording identity with compact metrics and a detail path.
*_events	Event in a recording view	Parsed step, message, tool call, observation, rejected response, or usage record. Use canonical_view for aggregate behavior.
sources	Source-file read	Original and redacted-copy hashes, parser version, file coordinates, and provenance.
annotations	Annotation	Existing judge interpretation, command analysis, generation gate, or recorded trial grade. evidence_layer identifies its scope.
examples	Derived example	Forest process/final/rejection export row or OLMo Experiment 1 packet reference. Nested source data remain JSON strings.
example_summary	Derived example	Compact metadata without exported messages, for the notebook's first load.
coverage	Missing/invalid source record or population entry	Explicit gaps and invalid sources. A missing outcome stays unknown.
forest_joins	Public example's source identity	Exact audit trial UUID and complete branch suffix, candidate count, and match status.
token_diagnostics	Recorded completion sequence	Selected-token count, selected-logprob count, alignment, summed negative logprob, and surprisal.
redactions	Redaction operation	Source path, JSON/text coordinate, and removal kind. No removed values are published.
q09_transitions	Transition	446 auxiliary training-log rows from one recorded attempt. These are not complete conversations.
qwen27b_summary_only	Existing summary record	64 Qwen 27B records with no local raw recording. These are not added to the episode population.
usage_events	Usage event in a recording view	Provider counters from completed Codex/OpenCode events. These remain separate from episode-level usage counters and retain view identity.

Episode bundles at recordings/<collection>/<episode_id>.json.gz contain recording views, redacted source objects, metadata, normalized events, and diagnostic coordinates. Metadata such as forest grades added in the enrichment pass is authoritative in the Parquet tables and annotation records; the bundle retains the original export summary. Invalid files are retained under invalid/, including redacted raw text and parser errors. Derived forest exports are compressed JSONL files under derived/forest/.

Sources and attribution

Local recordings come from the user's Olmo-RewardHacking, rl_experiments, and forestOfAudits workspaces. Source-file SHA-256 hashes, repository HEADs, dirty-state flags, checkpoint revisions where present, and original relative paths appear in provenance/ and the sources table. Dirty worktrees are reported because Git HEAD alone cannot identify their artifacts. The original files were read without alteration.

The public forest examples come from pranay5255/forestofaudits-agent-grader-traces-20260608-backfilled, pinned to revision b78e6f7adc3e14323d877d61edd94432938ddfe8. Its 8,633 process examples, 218 final-answer examples, and 74 rejection records preserve row order, metadata, truncation flags, and source coordinates. Its README, dataset summary, and audit split metadata are included under provenance/forest_public/. Of the process examples, 4,547 have the source truncated flag. They are partial exported contexts, not full trajectories. Gold/reference final answers may have no corresponding agent trajectory.

Forest records join only when the audit trial UUID and branch/file suffix identify one local episode. This verifies source identity, not equality of the exported, possibly truncated messages. Uncertain matches remain separate and flagged. Recorded forest grader results are trial-level verifier evidence; a trial grade is not assigned as an individual worker or judge branch outcome.

The 144 frozen OLMo Experiment 1 judge packets and existing GPT-5/Luna labels are derived evidence. The original think-instruct-pilot/20260822T114511Z raw run is missing locally. Existing Experiment 2 judge labels are retained as annotations. Their recovered label rows lack a verified episode identity, so they are not silently joined by position. Judge attempts and paired label comparisons remain distinct from recorded agent actions and verifier results.

Underlying benchmark prompts, audit excerpts, code, model output, and third-party export content retain their source attribution and applicable rights. This mixed corpus does not assert a new blanket license over those materials. The pinned forest source does not declare a dataset license in its card.

Coverage and measurement limits

• The 570 OLMo episodes include the retained 503 Experiment 2 identities and 67 preliminary/probe episodes. Native Mini-SWE companions can contain generations rejected by the parser, even when the accepted ATIF view has no agent actions.
• The Qwen archive has 512 files: 510 parseable recordings and two empty files. Its frozen remote index lists 880 candidate trial paths, of which 368 lacked a persisted trajectory at archive time. This is historical coverage, not a live Modal volume check.
• Qwen's 64 summary-only 27B records form a separate summary collection. Gate reward arrays keep their task/repetition grouping. They are never mapped to an episode by array position. Existing per-recording effective rewards keep their source-path join and selection/retry role.
• One malformed forest OpenCode stream is preserved as invalid coverage. Initial-only and empty recordings remain represented by their existing identities.
• Command categories are overlapping observable indicators. Chat code fences can describe proposed commands. Use the accompanying observations to establish execution. An evaluator-file reference, a file edit, a judge intent label, and a successful verifier exploit are different outcomes.
• Selected-token surprisal uses only aligned stored selected log probabilities. It is not model entropy or an alternative-action likelihood. Token-count alignment does not establish decoded transcript equality. OLMo telemetry sequence coordinates are not assumed to equal accepted action-turn numbers.
• Missing model revisions, limits, prices, grades, retries, and telemetry stay missing. Stored completion counters may include reasoning and rejected generations. Dollar cost is reported only when recorded; no current price assumptions are used.
• context_limit records the model limit only when the variant manifest states it. Qwen's sequence_limit retains the generation/training sequence cap separately. Recorded event usage remains separate because counters can have different accumulation conventions.
• Tasks are the aggregation and bootstrap unit. The collections do not support causal checkpoint claims or population-level safety claims. Forest branches sharing a trial are not independent trials.

The tmax collection and Codex development-session archives are excluded. Codex benchmark agents recorded inside forestOfAudits are included.

Redaction and reproducibility

Publication copies remove credential/authentication fields and recognizable credentials from text, environment settings, URL authentication, and command strings. Numeric token IDs and selected-logprob telemetry remain available. source_sha256 identifies the original bytes; publication_sha256 identifies the serialized redacted object or retained derived copy. JSON formatting can differ even when no secret was present. Redaction changes publication text lengths and may prevent exact text reconstruction from token telemetry.

Export and validation scripts are in scripts/. provenance/build_summary.json records counts and source revisions. provenance/validation.json records bundle hashes, coverage, joins, event counts, and token recomputation. CPU notebook execution reports appear in provenance/ after successful validation.

The full audit verified 1,451 episode bundles, 4,080 embedded source hashes, 64,252 normalized events, and 18,966 aligned token sequences. There are 5,910 rejected OLMo response records in the retained companion views. Exact forest source coordinates join 8,611 public example records to local episodes; 94 source-bearing records remain unmatched. These are example counts, not distinct trajectory counts.

The notebook passed in a fresh CPU kernel using about 253 MiB of memory locally and about 265 MiB when downloading anonymously from the public repository. The remote run completed in about 40 seconds, including inspector callbacks, a truncated process example, and a judge packet load. All 17 dataset configurations and public artifact hashes passed anonymous verification. Reports identify the tested dataset commit; later report-only commits keep the tested data unchanged.

To rebuild from the original workspace and the pinned public release:

python -m pip install -r requirements.txt
hf download pranay5255/forestofaudits-agent-grader-traces-20260608-backfilled --repo-type dataset --revision b78e6f7adc3e14323d877d61edd94432938ddfe8 --local-dir ../trajectory_study_inputs/forest_public
python scripts/export_corpus.py --root /home/experiments_base --forest-public ../trajectory_study_inputs/forest_public
python scripts/enrich_metadata.py
python scripts/make_notebook.py
python -m unittest discover -s scripts -p test_export.py -v
python scripts/validate_corpus.py --check-originals
python scripts/validate_metadata.py
python scripts/execute_notebook.py

The exporter consumes the audited trajectory_inventory/recordings.csv and datasets.csv from the original workspace. It verifies recording hashes and fails on input drift or a changed target population. Publication copies remain separate from those inputs.