Datasets:

ziyan98
/

E2K-cd82-harness-knowledge

Exception:    SplitsNotFoundError
Message:      The split names could not be parsed from the dataset config.
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 290, in _generate_tables
                  pa_table = paj.read_json(
                      io.BytesIO(batch), read_options=paj.ReadOptions(block_size=block_size)
                  )
                File "pyarrow/_json.pyx", line 364, in pyarrow._json.read_json
                File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
                  return check_status(status)
                File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
                  raise convert_status(status)
              pyarrow.lib.ArrowInvalid: JSON parse error: Column() changed from object to string in row 0
              
              During handling of the above exception, another exception occurred:
              
              Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
                  for split_generator in builder._split_generators(
                                         ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 101, in _split_generators
                  pa_table = next(iter(self._generate_tables(**splits[0].gen_kwargs, allow_full_read=False)))[1]
                             ~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 304, in _generate_tables
                  batch = json_encode_fields_in_json_lines(original_batch, json_field_paths)
                File "/usr/local/lib/python3.14/site-packages/datasets/utils/json.py", line 111, in json_encode_fields_in_json_lines
                  examples = [ujson_loads(line) for line in original_batch.splitlines()]
                              ~~~~~~~~~~~^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/utils/json.py", line 20, in ujson_loads
                  return pd.io.json.ujson_loads(*args, **kwargs)
                         ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
              ValueError: Expected object or value
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 68, in compute_split_names_from_streaming_response
                  for split in get_dataset_split_names(
                               ~~~~~~~~~~~~~~~~~~~~~~~^
                      path=dataset,
                      ^^^^^^^^^^^^^
                      config_name=config,
                      ^^^^^^^^^^^^^^^^^^^
                      token=hf_token,
                      ^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
                  info = get_dataset_config_info(
                      path,
                  ...<6 lines>...
                      **config_kwargs,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
                  raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
              datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.

Need 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 cd82 harness knowledge — the offline knowledge pipeline for ARC-AGI-3 game `cd82-fb555c5d`

What this is. Everything the E2K project produced while turning experience on one ARC-AGI-3 game (cd82-fb555c5d, case ARC3_004 of E2K-40) into verified, atomic knowledge between 2026-09-03 and 2026-09-12: 100-rollout play corpora per level, faceted claim extraction, two generations of claim verification with real probes, the knowledge re-injection experiment that closed the loop, and the final knowledge build — 888 atoms (knowledge_set.jsonl) and the 391-statement claim form (claim_form/knowledge_set.jsonl) that the RL arms in ziyan98/E2K-cd82-rl-internalization were actually trained from (see the rl_config_*.json files; the atom-set arms launched on 2026-09-12 were cancelled the same day). Code: ziyan-wang98/knowledge-learning-e2k (scripts/isambard/run_harness_e2k.sh, src/knowledge_learning/loop/{knowledge_filter,mechanism_registry,merge_levels,atomize}.py; private at the time of this release, available on request).

The harness. The game is served to an agent CLI as MCP stdio tools (loop/arc_mcp.py): look, act, status in play and extraction, plus set_level for verifier v2 to probe other levels. The actor is Qwen3.8-27B (vLLM, one node per worker, tp=4, max_model_len 262144, reply cap 32768 tokens) inside Qwen Code 0.22.3 for play, extraction and verification. In the play stage every file and shell tool is removed: the model can only learn the game by playing it. Each rollout is journaled by the MCP server (journal_NNN.jsonl: action, cells changed, the observation grid) — the authoritative record — next to the streamed CLI transcript (rollout_NNN.jsonl: the assistant's text, tool calls and tool results as streamed by Qwen Code; no separate thinking/reasoning blocks are recorded).

Headline results

experiment	result
π₀ baseline, level 1 (`runs/ARC3_004`, 100 rollouts)	9/100 cleared level 1 (all episodes: median 54 actions; the 9 clears took 32–79 actions, median 59)
π₁-full: the 3 extracted claims incl. the 5-move recipe prepended as `QWEN.md` (`runs/ARC3_004-pi1-full`)	100/100, median 5 steps
π₁-mech: recipe removed, mechanism claims only (`runs/ARC3_004-pi1-mech`)	33/91 (36 %), z = +4.5 vs π₀
play from levels 2, 3, 4 with a one-level cap (`runs/ARC3_004-L{2,3,4}`, 100 each)	no episode cleared its level; the corpora are what the per-level claims were extracted from
extractor v3 per level → verifier v2 (forward + reverse jobs)	474 claims (101 / 121 / 128 / 124 on levels 1–4), 436 with a ledger (union), 423 verified at least in part
final build `knowledge-v2/20260912-0920-final/`	888 atoms (pass A 1,162 atoms → merge 939 → gates 888 kept / 51 excluded, 16 of them near-duplicate collapses), mechanism registry `cd82-v1` (18 mechanisms, 394 verifier keys mapped, 17 unmapped); claim form 391 corrected statements (474 claims → 391 kept)

So knowledge extracted from the model's own play (91 failed and 9 successful episodes) transfers back into play: the composed plan is strongest (100 %), atomic mechanism claims also help (36 % vs 9 %).

path	files	size
`SYMLINKS.tsv`	1	1.6 MiB
`claim-pool-review`	35	2.5 MiB
`e2k-offline`	1	169.1 KiB
`excluded-runs`	5	113.6 KiB
`knowledge-pi1`	2	23.1 KiB
`knowledge-v2`	5967	2.0 GiB
`registry`	13	1.9 MiB
`runs/ARC3_004`	415	208.9 MiB
`runs/ARC3_004-ALL-verify-v2`	3	2.0 MiB
`runs/ARC3_004-L1-extract-v3`	247	31.0 MiB
`runs/ARC3_004-L1-verify-v2`	397	207.1 MiB
`runs/ARC3_004-L1-verify-v2-rev`	156	79.9 MiB
`runs/ARC3_004-L2`	310	184.7 MiB
`runs/ARC3_004-L2-extract-v3`	253	28.9 MiB
`runs/ARC3_004-L2-verify-v2`	446	218.0 MiB
`runs/ARC3_004-L2-verify-v2-rev`	201	97.2 MiB
`runs/ARC3_004-L3`	310	202.2 MiB
`runs/ARC3_004-L3-extract-v3`	254	30.0 MiB
`runs/ARC3_004-L3-verify-v2`	511	291.5 MiB
`runs/ARC3_004-L3-verify-v2-rev`	194	93.3 MiB
`runs/ARC3_004-L4`	310	177.6 MiB
`runs/ARC3_004-L4-extract-v3`	255	29.3 MiB
`runs/ARC3_004-L4-verify-v2`	493	277.3 MiB
`runs/ARC3_004-L4-verify-v2-rev`	179	87.3 MiB
`runs/ARC3_004-extract-pool`	244	55.7 MiB
`runs/ARC3_004-pi1-full`	311	18.9 MiB
`runs/ARC3_004-pi1-mech`	284	173.5 MiB
`runs/ARC3_004-round2`	9	228.3 KiB
`runs/ARC3_004-verify`	408	116.7 MiB
`runs/ARC3_004-verify-pilot`	17	2.4 MiB
`slurm-logs`	2	20.3 KiB
`state`	15	199 B
`tools`	1555	207.0 MiB
total	13803	4.8 GiB

`runs/<run>/` — one directory per harness job

run	stage	what is inside
`ARC3_004`	play, level 1, π₀	100 journals + rollouts, `extraction.jsonl`, `claims.json` (the 3 claims used for π₁), `corpus/`
`ARC3_004-L2`, `-L3`, `-L4`	play from level 2 / 3 / 4, `E2K_LEVEL_CAP=1`	100 journals + rollouts each
`ARC3_004-pi1-full`, `-pi1-mech`	play with knowledge re-injected (`QWEN.md` is the injected block)	100 and 91 journals
`ARC3_004-extract-pool`	faceted extraction over the level-1 corpus	35 `claims_NNN.json` passes → `pool.json` (105 claims)
`ARC3_004-verify-pilot`	verifier v1 pilot over the 3 π₁ claims (2026-09-07)	`claims_under_test.json`, 3 ledgers, 3 verify journals
`ARC3_004-verify`	verifier v1 over the 105-claim pool	`claims_under_test.json`, 81 `ledger_NNN.json`, 102 verify journals, `verdicts.json`, `filtered/knowledge_set.jsonl` (61 kept — audited, see `claim-pool-review/`)
`ARC3_004-L{1..4}-extract-v3`	extractor v3 per level	36–44 `claims_NNN.json` per level, `level_conflicts.json`, `duplicate_candidates.json`, `corpus/`
`ARC3_004-L{1..4}-verify-v2`, `-rev`	verifier v2, each level from both ends of the claim list	`claims_under_test.json`, ledgers (83+29, 76+35, 116+37, 110+30), verify journals
`ARC3_004-ALL-verify-v2`	a `merge_levels` snapshot of the four forward verify-v2 runs taken 2026-09-11 14:43 while they were still running: 474 claims in one global index space, 178 ledgers / 267 verify journals linked at that moment; no verifier job ran here	`claims.json`, `merge_manifest.json` (its ledger/journal/log entries were symlinks into the per-level runs and are listed in `SYMLINKS.tsv`)
`ARC3_004-round2`	a 2026-09-08 harness round that left only workspaces and logs	—

The journals/ corpus each verifier read was a directory of symlinks to the play run's journal_NNN.jsonl (runs/ARC3_004, -L2, -L3, -L4); it is not materialised here (see SYMLINKS.tsv), the same for the extract runs' journal/rollout links. Every run also has logs/ (per-rollout CLI logs, worker preflights) and workspace_and_sessions.tar.zst: the per-rollout working directories (workspace/w*/run_NNN) and the Qwen Code session logs (qwen-home/projects). Compile caches, core dumps and extension stores were dropped.

`knowledge-v2/` — the knowledge-set builds

Three snapshots of scripts/isambard/finalize_knowledge_e2k.sh. 20260911-1653 and 20260911-1659 are interim filter-only builds (union, merged/, filter/ with --dedup-mode apply; their knowledge_set.jsonl is the clause-level filter output, 103 and 18 rows; 1653 has no registry). 20260912-0920-final is the release build: per-level verification union (levels/L{1..4}-verify/), merged/ (one global index space) + merged/mechanisms.json (registry), filter/ (clause-level knowledge_filter --dedup-mode propose output), atoms/ (the two model passes of loop.atomize with per-claim caches), knowledge_set.jsonl + .sha256 (888 atoms), atoms_report.{md,json}, SUMMARY.md, mechanism_registry_report.md, claim_form/ (the 391 corrected claim statements the RL arms trained on), recall_forms_v1/v2 (the sentence forms carried by the IDD teacher and the arm-C rows, and used for recital metrics), and the 20 rl_config_*.json files the RL arms were launched with.

Other directories

knowledge-pi1/ — claims_full.json (3 claims) and claims_mech.json (2 claims): the blocks injected in π₁-full and π₁-mech.
registry/ — the mechanism registry build and the filter-with/without-registry summaries. The registry gives each of its 18 mechanisms the number of the matching hand-written oracle rule and a one-paragraph author statement of the mechanism; it was used only to map verifier keys and was never shown to the extractor or verifier.
tools/ — atoms_v1 (first atomize build), filter previews per level, the chain Slurm scripts and finish_proof.
claim-pool-review/ — the deep review of the 105-claim pool, performed by model agents (two Claude Code workflow passes, 11 reviewer chunks) against the game source: 77 reviewed; of the 61 kept by filter v1 only 4 survive as written, 19 are click recipes, 28 are partly refuted. synthesis/ holds the resulting verifier-v2 / extractor-v3 prompt changes (its SUMMARY.md predates that directory and still says those artefacts do not exist).
e2k-offline/ — the MCP action journal (17 act calls with their results) of the 2026-09-03 scripted offline probe of the game, recorded before the first harness run.
state/ — the chain state files (job ids per level).
excluded-runs/ — logs only of five early harness attempts of 2026-09-03 that failed on infrastructure (context window, tool parser, 16K truncation) or were invalid (the model read the rules file); their play is not data.
slurm-logs/ — harness, atomize and link-play job logs.

Caveats

Claims are hypotheses the verifier judged. No verdict was audited by a human; the only independent check is the model-agent pool review above, which covers 77 of the 105 v1-pool claims and none of the 474 v2 claims behind the 888 atoms.
The ARC-AGI-3 game source (cd82.py, MIT, ARC Prize Foundation) is not part of this release and was never shown to the extractor or verifier; the hand-written oracle-rules file is not included either (the registry carries only its rule numbers and the author's mechanism statements).

Integrity and licence

SYMLINKS.tsv lists every symbolic link of the source tree (6,741 rows: the 3,999 links of the three merged/ directories, whose targets live in levels/, plus 2,742 links inside the run directories — the play journals/rollouts linked into the extract and verify runs, and ALL-verify-v2's ledger/journal/log links; run-level link paths are given relative to their run directory). MANIFEST.sha256 / MANIFEST.json hash every data file (the README and the manifests themselves are not listed). CC BY 4.0 (ARC-AGI-3 observations under their own terms). Ziyan Wang, King's College London, 2026; compute: Isambard-AI (BriCS).