E2K: experience to knowledge
Collection
Entry point: 5 datasets (trajectories, verified knowledge, RL runs) + one checkpoint per RL run; each run has its own series collection. • 24 items • Updated
Dataset Viewer
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
validated: bool
required_validation_gates_passed: bool
case_id: string
validation_phase: string
trajectory: struct<attempt-01: struct<attempt_id: string, decision_files: int64, result_files: int64, paired_cou (... 1872 chars omitted)
child 0, attempt-01: struct<attempt_id: string, decision_files: int64, result_files: int64, paired_counts_match: bool, tu (... 712 chars omitted)
child 0, attempt_id: string
child 1, decision_files: int64
child 2, result_files: int64
child 3, paired_counts_match: bool
child 4, turn_numbers_contiguous: bool
child 5, step_indices_match_turns: bool
child 6, decision_and_result_event_ids_match_filenames: bool
child 7, observable_state_chain_matches: bool
child 8, chosen_actions_were_legal: bool
child 9, session_id_is_constant: bool
child 10, inherited_snapshot_hash_in_every_decision: bool
child 11, blind_failure_marker_hash_in_every_post_failure_decision: bool
child 12, decision_phase_counts: struct<blind-independent: int64>
child 0, blind-independent: int64
child 13, local_claim_references: struct<>
child 14, inherited_claim_references: struct<>
child 15, level_completion_turns: list<item: int64>
child 0, item: int64
child 16, life_loss_turns: list<item: int64>
child 0, item: int64
child 17, game_won_turns: list<item: null>
child 0, item: null
child 18, final_result_step_index: int64
child
...
mpt-01_evidence_files_mode_0444: bool
child 3, oracle_or_source_inspection_before_failure_freeze: bool
child 4, failure_frozen_before_post_failure_session_start: bool
child 5, oracle_phase_entered_after_failure_freeze: bool
child 6, post_failure_sources_inspected: list<item: string>
child 0, item: string
child 7, authoritative_answer_or_game_rule_source_found: bool
child 8, empirical_probe_evidence_used_for_solution: bool
hash_audit: struct<attempt-01_sorted_turn_hashes: string, attempt-02_sorted_turn_hashes: string, combined_sorted (... 223 chars omitted)
child 0, attempt-01_sorted_turn_hashes: string
child 1, attempt-02_sorted_turn_hashes: string
child 2, combined_sorted_turn_hashes: string
child 3, inherited_claims_jsonl: string
child 4, inherited_knowledge_json: string
child 5, terminal_failure_marker: string
child 6, planned_atomic_skills_jsonl: string
child 7, planned_run_summary_json: string
child 8, collection_hash_method: string
secret_token_present_in_output: bool
token_leak_scan_file_count_before_final_artifacts: int64
token_value_or_hash_recorded: bool
json_artifacts_parse: bool
repository_worktree_clean: bool
repository_files_modified: bool
formal_e2k40_data_modified: bool
other_case_files_modified: bool
warnings: list<item: null>
child 0, item: null
errors: list<item: null>
child 0, item: null
oracle_consulted: bool
attempt_id: string
benchmark_or_client_source_inspected: bool
authoritative_answer_or_source_inspected: bool
to
{'validated': Value('bool'), 'required_validation_gates_passed': Value('bool'), 'case_id': Value('string'), 'attempt_id': Value('string'), 'validation_phase': Value('string'), 'trajectory': {'server_step_index_actions_applied': Value('int64'), 'trace_rows': Value('int64'), 'trace_turn_range': List(Value('int64')), 'trace_turn_numbers_complete': Value('bool'), 'trace_missing_turns': List(Value('int64')), 'trace_duplicate_turns': List(Value('null')), 'trace_gap_was_present_before_terminal_continuation': Value('bool'), 'retained_individual_decision_files': Value('int64'), 'retained_individual_result_files': Value('int64'), 'retained_individual_turn_range': List(Value('int64')), 'paired_counts_match': Value('bool'), 'retained_turn_numbers_contiguous': Value('bool'), 'step_indices_match_turns': Value('bool'), 'decision_turns_and_trace_event_ids_match_filenames': Value('bool'), 'observable_state_chain_matches_for_retained_pairs': Value('bool'), 'chosen_actions_were_legal_for_retained_pairs': Value('bool'), 'session_id_is_constant_for_retained_results': Value('bool'), 'level_boundary_transition_field_clears_on_next_request': List(Value('int64')), 'final_result_step_index': Value('int64'), 'final_status': Value('string'), 'final_done': Value('bool'), 'game_won_event_present': Value('bool'), 'levels_completed': Value('int64'), 'level_completion_turns': List(Value('int64')), 'lives_lost': Value('int64'), 'life_loss_turns': List(Value('null')), 'lives_left': Value('int64')}, 'atomic_claims': {'count': Value('int64'), 'exact_schema_and_field_order_valid': Value('bool'), 'unique_claim_ids': Value('bool'), 'evidence_ids_resolve_to_supplied_trace_events': Value('bool'), 'status_counts': {'supported': Value('int64'), 'contradicted': Value('int64'), 'hypothesis': Value('int64')}, 'provenance_counts': {'blind': Value('int64'), 'post_failure_oracle': Value('int64')}}, 'inherited_snapshot': {'claim_count': Value('int64'), 'manifest_sha256': Value('string'), 'read_before_first_interaction': Value('bool'), 'p2_claims_assumed_without_p2_evidence': Value('bool')}, 'json_artifacts_parse': Value('bool'), 'secret_token_present_in_output': Value('bool'), 'oracle_consulted': Value('bool'), 'authoritative_answer_or_source_inspected': Value('bool'), 'benchmark_or_client_source_inspected': Value('bool'), 'formal_e2k40_data_modified': Value('bool'), 'warnings': List(Value('string')), 'errors': List(Value('null'))}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2951, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 547, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
validated: bool
required_validation_gates_passed: bool
case_id: string
validation_phase: string
trajectory: struct<attempt-01: struct<attempt_id: string, decision_files: int64, result_files: int64, paired_cou (... 1872 chars omitted)
child 0, attempt-01: struct<attempt_id: string, decision_files: int64, result_files: int64, paired_counts_match: bool, tu (... 712 chars omitted)
child 0, attempt_id: string
child 1, decision_files: int64
child 2, result_files: int64
child 3, paired_counts_match: bool
child 4, turn_numbers_contiguous: bool
child 5, step_indices_match_turns: bool
child 6, decision_and_result_event_ids_match_filenames: bool
child 7, observable_state_chain_matches: bool
child 8, chosen_actions_were_legal: bool
child 9, session_id_is_constant: bool
child 10, inherited_snapshot_hash_in_every_decision: bool
child 11, blind_failure_marker_hash_in_every_post_failure_decision: bool
child 12, decision_phase_counts: struct<blind-independent: int64>
child 0, blind-independent: int64
child 13, local_claim_references: struct<>
child 14, inherited_claim_references: struct<>
child 15, level_completion_turns: list<item: int64>
child 0, item: int64
child 16, life_loss_turns: list<item: int64>
child 0, item: int64
child 17, game_won_turns: list<item: null>
child 0, item: null
child 18, final_result_step_index: int64
child
...
mpt-01_evidence_files_mode_0444: bool
child 3, oracle_or_source_inspection_before_failure_freeze: bool
child 4, failure_frozen_before_post_failure_session_start: bool
child 5, oracle_phase_entered_after_failure_freeze: bool
child 6, post_failure_sources_inspected: list<item: string>
child 0, item: string
child 7, authoritative_answer_or_game_rule_source_found: bool
child 8, empirical_probe_evidence_used_for_solution: bool
hash_audit: struct<attempt-01_sorted_turn_hashes: string, attempt-02_sorted_turn_hashes: string, combined_sorted (... 223 chars omitted)
child 0, attempt-01_sorted_turn_hashes: string
child 1, attempt-02_sorted_turn_hashes: string
child 2, combined_sorted_turn_hashes: string
child 3, inherited_claims_jsonl: string
child 4, inherited_knowledge_json: string
child 5, terminal_failure_marker: string
child 6, planned_atomic_skills_jsonl: string
child 7, planned_run_summary_json: string
child 8, collection_hash_method: string
secret_token_present_in_output: bool
token_leak_scan_file_count_before_final_artifacts: int64
token_value_or_hash_recorded: bool
json_artifacts_parse: bool
repository_worktree_clean: bool
repository_files_modified: bool
formal_e2k40_data_modified: bool
other_case_files_modified: bool
warnings: list<item: null>
child 0, item: null
errors: list<item: null>
child 0, item: null
oracle_consulted: bool
attempt_id: string
benchmark_or_client_source_inspected: bool
authoritative_answer_or_source_inspected: bool
to
{'validated': Value('bool'), 'required_validation_gates_passed': Value('bool'), 'case_id': Value('string'), 'attempt_id': Value('string'), 'validation_phase': Value('string'), 'trajectory': {'server_step_index_actions_applied': Value('int64'), 'trace_rows': Value('int64'), 'trace_turn_range': List(Value('int64')), 'trace_turn_numbers_complete': Value('bool'), 'trace_missing_turns': List(Value('int64')), 'trace_duplicate_turns': List(Value('null')), 'trace_gap_was_present_before_terminal_continuation': Value('bool'), 'retained_individual_decision_files': Value('int64'), 'retained_individual_result_files': Value('int64'), 'retained_individual_turn_range': List(Value('int64')), 'paired_counts_match': Value('bool'), 'retained_turn_numbers_contiguous': Value('bool'), 'step_indices_match_turns': Value('bool'), 'decision_turns_and_trace_event_ids_match_filenames': Value('bool'), 'observable_state_chain_matches_for_retained_pairs': Value('bool'), 'chosen_actions_were_legal_for_retained_pairs': Value('bool'), 'session_id_is_constant_for_retained_results': Value('bool'), 'level_boundary_transition_field_clears_on_next_request': List(Value('int64')), 'final_result_step_index': Value('int64'), 'final_status': Value('string'), 'final_done': Value('bool'), 'game_won_event_present': Value('bool'), 'levels_completed': Value('int64'), 'level_completion_turns': List(Value('int64')), 'lives_lost': Value('int64'), 'life_loss_turns': List(Value('null')), 'lives_left': Value('int64')}, 'atomic_claims': {'count': Value('int64'), 'exact_schema_and_field_order_valid': Value('bool'), 'unique_claim_ids': Value('bool'), 'evidence_ids_resolve_to_supplied_trace_events': Value('bool'), 'status_counts': {'supported': Value('int64'), 'contradicted': Value('int64'), 'hypothesis': Value('int64')}, 'provenance_counts': {'blind': Value('int64'), 'post_failure_oracle': Value('int64')}}, 'inherited_snapshot': {'claim_count': Value('int64'), 'manifest_sha256': Value('string'), 'read_before_first_interaction': Value('bool'), 'p2_claims_assumed_without_p2_evidence': Value('bool')}, 'json_artifacts_parse': Value('bool'), 'secret_token_present_in_output': Value('bool'), 'oracle_consulted': Value('bool'), 'authoritative_answer_or_source_inspected': Value('bool'), 'benchmark_or_client_source_inspected': Value('bool'), 'formal_e2k40_data_modified': Value('bool'), 'warnings': List(Value('string')), 'errors': List(Value('null'))}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
E2K-40 v0.1 — experience-to-knowledge measurement trajectories
What this is. The frozen layer-1 input of the E2K (experience → knowledge) project: 40
interaction trajectories, 20 on ARC-AGI-3 games and 20 on DiG-bench puzzles, collected on
2026-08-30 with a Qwen3.8-27B actor, projected into reasoning-free bundles that an external
knowledge extractor is allowed to see, plus the collector attempts, container receipts and
Slurm logs behind them. The repository that produced it is
ziyan-wang98/knowledge-learning-e2k (E2K-40 v0.1 pipeline, src/knowledge_learning/e2k/;
private at the time of this release, available on request).
Status of the measurement protocol. Collection and ingest are complete (40/40 main cases,
reports/validation.md = draft gate PASS). The second half of the v0.1 protocol — sending each
bundle to GPT-5.6 Pro for one knowledge claim and labelling it on four axes (truth, trace support,
atomicity, scope) — was not carried out: extraction_status is empty on every row and every
main case's annotation_status is pending (the P-21 holdout row is not_applicable), and
dataset.json still says input_state: open. The project moved on to self-extraction with the same
actor (see ziyan98/E2K-loop-v0.2) and
then to a single-game knowledge pipeline
(ziyan98/E2K-cd82-harness-knowledge).
Contents
| path | files | size |
|---|---|---|
annotation-ui |
2 | 66 B |
collector-runs |
274 | 222.1 MiB |
discovery |
713 | 17.5 MiB |
dry-runs |
291 | 17.7 MiB |
e2k40-v0.1 |
1993 | 202.2 MiB |
slurm-logs |
2 | 976.3 KiB |
| total | 3275 | 460.4 MiB |
e2k40-v0.1/ — the dataset root (formal run)
ARC-AGI-3 (arc_agi3/ARC3_001..020) |
DiG-bench (dig_bench/DIG_001..020) |
|
|---|---|---|
| cases | 20 games (manifest.csv column source_game_id) |
20 puzzles P-1 … P-20 (P-21 is a held-out row with no artifacts) |
| trajectory outcome | 8 failure (GAME_OVER), 12 progress (the 100-decision budget ran out without WIN or GAME_OVER; 10 of the 12 with 0 levels completed) |
8 failure (no level beaten: 6 at the 100-turn budget, 2 terminal game over), 12 progress (1–4 levels beaten when the budget ran out) |
| transitions | 1,314 | 1,933 |
| actor | Qwen3.8-27B via vLLM 0.27.1, sampling contract v1.1 (max_tokens 4096, 4 retries per action, 100-decision budget, fixed seed) |
same, through the pinned dig-bench harness (commit d88405f) with an audited after state |
The two outcome labels are assigned by different rules (terminal state for ARC, levels beaten for DiG), so they are not comparable across the two environments.
Per case:
raw_recording.jsonl— immutable evidence; every actor call including its reasoning.trace.json/trace.md— allow-listed projection with stableevent_ids (no reasoning, no labels, no source ids).metadata.json; for ARC cases alsoframes/*.png(byte-deterministic renders) andcontact_sheet_*.png. DiG cases are text-only.model_bundle/—web_input.mdandinput_manifest.json, plus the contact sheets for ARC cases: the only material an extractor may see; the blind ids areVISUAL_CASE_NNN/TEXT_CASE_NNN.actor_model_calls.jsonl(ARC) — the request/response log of the collector.
Root files: manifest.csv (one row per case, sha256 of every input), selection.json
(seed 20260824), extraction_queue.csv, configs/e2k40-v0.1.json (pins the extractor prompt
sha256, the ARC candidate universe and both upstream commits), prompts/extractor_v1.txt,
reports/ (validation and empty annotation summaries), README.md (the workspace note written
by init) and an empty logs/. The root also carries _smoke/manual-70018-0/, a copy of the
sealed synthetic smoke fixture (run_type: synthetic_demo, same content as
dry-runs/e2k40-smoke-manual-20260825/): its gpt56pro_raw.md, extracted_claim.json and
annotation.json files are generated test data, not GPT output or human labels. The scientific
protocol and the metric definitions are docs/E2K40_PROTOCOL.md and
docs/ANNOTATION_MEASUREMENTS.md in the code repository.
dry-runs/
e2k40-dry-run-v0.1/— the 2 ARC + 2 DiG dry run under contract v1.1 (2026-08-30), the run that fixed the contract before the formal collection.e2k40-dry-run-v0.1.contract2048-20260830T0442Z/— the archived, partial dry run under the first contract (max_tokens2048, 2 retries): only ARC3_001 was collected and ingested; ARC3_002 died after three consecutive length-truncated turns (its attempts are undercollector-runs/dry-run.contract2048-20260830T0442Z/) and the DiG cases never ran. 19–28 % of ARC turns exhausted the 2,048-token budget inside the thinking block, which is why v1.1 exists.e2k40-smoke-manual-20260825/— a synthetic, sealed demo fixture (run_type: synthetic_demo); not research data.
collector-runs/
For every case the collector's attempts and the chosen_attempt.json that names the attempt whose
recording was ingested. ARC attempts hold attempt_NNN/collection.json, model_calls.jsonl and
official_recordings/ (the chosen-attempt file records the sha256 of recording, calls and
collection); DiG attempts hold collector_receipt.json, the harness recording
<run_label>_<timestamp>_<session>.jsonl and its .log (the chosen-attempt file records the sha256
of recording and receipt). _runtime/vllm-<job>.json are vLLM job receipts (container digest, model
config sha256, sampling contract, and the shard assignment for shard-mode jobs): 72 files, of which
the 12 collection.mode: shard receipts of 2026-08-30 (jobs 6200661, 6200697–6200703, 6201334,
6201771, 6203008) belong to the formal collection; vllm-6159544.json is the aborted 2026-08-28
attempt under the 2048-token contract, and the remaining 59 (extract, teach) are receipts of
the later v0.2 loop jobs that reused the same output root.
discovery/codex-sol-ultra-loop/
An exploratory pass from 2026-08-25/26, before the formal collection, in which a code agent
(the OpenAI Codex CLI; model label recorded as gpt-5.6-sol) played 15 of the ARC games and 10
DiG puzzles blind, recorded events, decisions, frames and official recordings, and kept
"atomic skill" ledgers (atomic_skills.jsonl, inherited_*). Each attempt directory
(blind_attempt/, attempt-NN/, checkpoints/) is packed as one .tar.zst archive next to its
case-level ledgers (26 archives; unpack with zstd -dc X.tar.zst | tar -x). dashboard/assets/
holds the renders used by the read-only discovery dashboard of the code repository. It is included
because the discovery dashboard reads it; it was not used by any later stage.
annotation-ui/
The access log (one request, 2026-08-25) and PID file of the local annotation UI that was started once against this root; no annotations were made. Kept only for completeness.
slurm-logs/
collection-slurm-logs.tar.zst: the collector, ingest and image-build job logs plus every vLLM
server log under logs/e2k40/vllm-*.log — the collection shards and dry runs, and also the 57
teachability and 3 self-extraction servers of the later v0.2 loop, which were archived together
(176 files; the list is in the sidecar .filelist.txt). Node-hour accounting of the whole project is
in the code repository.
Not included, and why
- The ARC-AGI-3 environment files (game source; MIT-licensed by the ARC Prize Foundation, fetched by
arc-agi0.9.9 /arcengine0.9.3) and the dig-bench harness checkout: obtainable upstream, and the extractor must never see them. - Container images (
vllm-openai-0.27.1-aarch64-cu129.sif): rebuildable fromcontainers/*.defin the code repository.
Integrity
MANIFEST.sha256 lists the sha256 of every data file in this repository (3,275 files; this
README and the two manifest files themselves are not listed); MANIFEST.json adds sizes.
SYMLINKS.tsv (when present) records symbolic links of the source tree that were not copied.
No file was altered on the way in, except that the discovery attempt directories were packed into
.tar.zst archives; the dataset-root files keep the hashes recorded in manifest.csv.
Citation
Ziyan Wang (King's College London), E2K: experience-to-knowledge measurement, 2026. Data collected on Isambard-AI (BriCS, University of Bristol). Third-party content (ARC-AGI-3 observations, DiG-bench puzzles) remains under its own terms; the trajectories and bundles are released under CC BY 4.0.
- Downloads last month
- 215