The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
type: string
subtype: string
uuid: string
session_id: string
cwd: string
tools: list<item: string>
child 0, item: string
mcp_servers: list<item: struct<name: string, status: string>>
child 0, item: struct<name: string, status: string>
child 0, name: string
child 1, status: string
model: string
permission_mode: string
slash_commands: list<item: string>
child 0, item: string
qwen_code_version: string
agents: list<item: string>
child 0, item: string
result: null
parent_tool_use_id: null
event: struct<type: string, goal_state: struct<v: int64, goal: null, activity: string>>
child 0, type: string
child 1, goal_state: struct<v: int64, goal: null, activity: string>
child 0, v: int64
child 1, goal: null
child 2, activity: string
message: struct<id: string, type: string, role: string, model: string, content: list<item: struct<type: strin (... 298 chars omitted)
child 0, id: string
child 1, type: string
child 2, role: string
child 3, model: string
child 4, content: list<item: struct<type: string, thinking: string, signature: string, text: string, id: string, name: (... 94 chars omitted)
child 0, item: struct<type: string, thinking: string, signature: string, text: string, id: string, name: string, in (... 82 chars omitted)
child 0, type: string
child 1, thinking: string
child 2, signature: string
child 3, text: string
child 4, id: string
child 5, name: string
child 6, input: struct<action: string>
child 0, action: string
child 7, tool_use_id: string
child 8, is_error: bool
child 9, content: string
child 5, stop_reason: string
child 6, usage: struct<input_tokens: int64, output_tokens: int64, cache_read_input_tokens: int64, total_tokens: int6 (... 2 chars omitted)
child 0, input_tokens: int64
child 1, output_tokens: int64
child 2, cache_read_input_tokens: int64
child 3, total_tokens: int64
arguments: struct<action: string, x: null, y: null>
child 0, action: string
child 1, x: null
child 2, y: null
tool: string
to
{'arguments': {'action': Value('string'), 'x': Value('null'), 'y': Value('null')}, 'result': {'actions': List({'action_type': Value('string'), 'requires_coordinates': Value('bool')}), 'actions_left': Value('int64'), 'cells_changed': Value('int64'), 'change_kinds': Json(decode=True), 'changed_cells': List(List(Value('int64'))), 'changed_omitted': Value('int64'), 'decisions': Value('int64'), 'done': Value('bool'), 'observation': Value('string'), 'state': Value('string')}, 'tool': Value('string')}
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
type: string
subtype: string
uuid: string
session_id: string
cwd: string
tools: list<item: string>
child 0, item: string
mcp_servers: list<item: struct<name: string, status: string>>
child 0, item: struct<name: string, status: string>
child 0, name: string
child 1, status: string
model: string
permission_mode: string
slash_commands: list<item: string>
child 0, item: string
qwen_code_version: string
agents: list<item: string>
child 0, item: string
result: null
parent_tool_use_id: null
event: struct<type: string, goal_state: struct<v: int64, goal: null, activity: string>>
child 0, type: string
child 1, goal_state: struct<v: int64, goal: null, activity: string>
child 0, v: int64
child 1, goal: null
child 2, activity: string
message: struct<id: string, type: string, role: string, model: string, content: list<item: struct<type: strin (... 298 chars omitted)
child 0, id: string
child 1, type: string
child 2, role: string
child 3, model: string
child 4, content: list<item: struct<type: string, thinking: string, signature: string, text: string, id: string, name: (... 94 chars omitted)
child 0, item: struct<type: string, thinking: string, signature: string, text: string, id: string, name: string, in (... 82 chars omitted)
child 0, type: string
child 1, thinking: string
child 2, signature: string
child 3, text: string
child 4, id: string
child 5, name: string
child 6, input: struct<action: string>
child 0, action: string
child 7, tool_use_id: string
child 8, is_error: bool
child 9, content: string
child 5, stop_reason: string
child 6, usage: struct<input_tokens: int64, output_tokens: int64, cache_read_input_tokens: int64, total_tokens: int6 (... 2 chars omitted)
child 0, input_tokens: int64
child 1, output_tokens: int64
child 2, cache_read_input_tokens: int64
child 3, total_tokens: int64
arguments: struct<action: string, x: null, y: null>
child 0, action: string
child 1, x: null
child 2, y: null
tool: string
to
{'arguments': {'action': Value('string'), 'x': Value('null'), 'y': Value('null')}, 'result': {'actions': List({'action_type': Value('string'), 'requires_coordinates': Value('bool')}), 'actions_left': Value('int64'), 'cells_changed': Value('int64'), 'change_kinds': Json(decode=True), 'changed_cells': List(List(Value('int64'))), 'changed_omitted': Value('int64'), 'decisions': Value('int64'), 'done': Value('bool'), 'observation': Value('string'), 'state': Value('string')}, 'tool': Value('string')}
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 cd82 RL internalization — every run of the knowledge-SFT + RL experiments
What this is. The complete record of the experiments that tried to write the verified
knowledge of ARC-AGI-3 game cd82-fb555c5d (the 888-atom set of
ziyan98/E2K-cd82-harness-knowledge)
into the weights of Qwen3.8-27B — supervised rows only, never the RL prompt — and then improve
play with Dr.GRPO on real multi-step episodes: 2026-09-09 … 2026-09-19 on Isambard-AI
(4 × GH200 nodes per job, slime v0.3.2 = Megatron-LM TP4 × PP4 + 16 SGLang engines). Per run:
the harness episodes (every turn of every episode), eval dumps, the exact training batches, the
W&B offline logs (also exported to JSONL), the per-rollout play metrics parsed from the job logs,
the SFT row sets, the run log with every decision, and an index of all 84 Megatron checkpoint
iterations. Ten of those checkpoints are converted to Hugging Face format and published as
ziyan98/E2K-cd82-Qwen3.8-27B-* model repositories (listed below). Code:
ziyan-wang98/knowledge-learning-e2k (src/knowledge_learning/rl/,
scripts/isambard/run_rl_internalize_e2k.sh, docs/RL_KNOWLEDGE_SFT.md,
docs/E2K_RL_INTERNALIZATION.md).
Design in one paragraph
Each RL rollout plays 32 episodes of the game through the Claude Code harness (CLI 2.1.270,
model served by slime's SGLang engines as slime-actor; the same look / act / status MCP
tools as the knowledge harness; levels 1–4 for training with level_cap 1, a shaped return on top
of level clears) and takes 8 optimizer steps (global batch 4, one training sample per LLM call,
Dr.GRPO: per-token loss, no std normalisation, lr 5e-6). Knowledge enters only through SFT rows
on a schedule — warmup SFT-only rollouts first (sft_first), interleaved (sft_interleaved),
or mixed into every RL batch (sft_joint) — with three row sources: declarative chat rows built
from the knowledge set, harness rows (recorded harness turns, pre-tokenised: recite rows
written by a teacher that had the knowledge in context, actctrl rows of the base model's own
turns, replay rows), and the arm-C rows (knowledge sentences inserted at the top of full-length
base-model thinks, with a cite weight on the knowledge tokens). Adam moments are reset at every
SFT → RL boundary.
What the arms showed (first RL rollout after SFT = the SFT-only policy; 32 episodes, noise ≈ 0.05)
| rows used for SFT | shaped return | clears |
|---|---|---|
none (base model, three baseline-cc starts) |
−0.16 / −0.20 / −0.26 | 0 / 96 |
| act-ctrl: the base model's own harness turns, no knowledge (two runs) | −0.42 / −0.52 | 0 / 64 |
| IDD-recite: teacher turns written with the knowledge in the prompt (two runs) | −0.15 / +0.01 | 1 / 64 |
| arm C, 6 SFT rollouts, cite weight up to 32 (knowledge NLL 0.42/token) | +0.01 | 1 / 32 |
| declarative chat rows, tool-call ending (trial 4) | −0.16 | 0 / 32 |
Under RL the pure baseline reaches +0.08 and 2/32 clears at rollout 10; the IDD arm climbs
+0.01, +0.08, +0.08, +0.21, +0.19, +0.10, +0.25 over its seven RL rollouts. No arm clears more
than 1–2 of 32 episodes. Arm C with knowledge NLL 0.42 is the first policy that states knowledge
verbatim while playing (14.9 % of thinks, 80 distinct sentences); only 5 of 168 recited sentences
named the action being taken, so reciting is not yet using. Declarative chat rows that end with
a text answer collapse play into single-call episodes (trials 1–3); a tool-call ending removes the
collapse but shortens the thinks to a third. The full narrative, with job ids and timestamps, is
state/early_recital_snapshot_20260916.md; the numbers per rollout are in metrics/.
Contents
| path | files | size |
|---|---|---|
checkpoints |
4 | 55.8 KiB |
knowledge-set |
52 | 5.1 MiB |
metrics |
4 | 590.6 KiB |
numeric |
1509 | 103.5 MiB |
runs/act_ctrl |
7 | 1.9 MiB |
runs/act_ctrl-cc |
78 | 1.2 GiB |
runs/baseline |
36 | 22.6 MiB |
runs/baseline-cc |
120 | 2.8 GiB |
runs/baseline-cc-fail-6536263 |
35 | 13.6 MiB |
runs/baseline-cc-fail-6545990 |
35 | 13.1 MiB |
runs/baseline-cc-fail-6547610 |
34 | 6.8 MiB |
runs/baseline-cc-old-20260914 |
59 | 38.2 MiB |
runs/baseline-memrepro-A |
34 | 3.5 MiB |
runs/baseline-memrepro-C |
34 | 3.5 MiB |
runs/baseline-qc |
8 | 3.9 MiB |
runs/baseline-qcshake |
248 | 39.4 MiB |
runs/baseline-shaped |
8 | 2.8 MiB |
runs/baseline-v2 |
19 | 3.4 MiB |
runs/cd82-play-v1-fail-6546650 |
15 | 4.6 MiB |
runs/cd82-play-v2 |
46 | 137.1 MiB |
runs/cd82-play-v3 |
28 | 145.5 MiB |
runs/cd82-play-v4 |
29 | 115.5 MiB |
runs/cd82-play-v4a |
13 | 408.1 MiB |
runs/cd82-play-v4b |
13 | 453.5 MiB |
runs/config_curatedL1.json |
1 | 6.9 KiB |
runs/curatedL1_check |
2 | 87.6 KiB |
runs/dbg-forge-default |
24 | 660.2 KiB |
runs/dbg-forge-norope |
24 | 667.0 KiB |
runs/dbg-forward |
24 | 568.5 KiB |
runs/dbg-forward-norope |
24 | 571.4 KiB |
runs/dbg-probe |
24 | 644.5 KiB |
runs/dbg-probe-disk |
24 | 516.3 KiB |
runs/dbg-probe-nccl |
24 | 649.0 KiB |
runs/mask_knowledge |
96 | 1.1 MiB |
runs/memrepro |
4 | 3.5 MiB |
runs/sft_first |
24 | 4.7 MiB |
runs/sft_first-cc |
71 | 6.8 MiB |
runs/sft_first-claimform |
53 | 40.1 MiB |
runs/sft_first-curatedL1 |
24 | 2.4 MiB |
runs/sft_first-v2 |
19 | 3.4 MiB |
runs/sft_first-v3 |
8 | 4.1 MiB |
runs/sft_first_armc-cc |
71 | 315.8 MiB |
runs/sft_first_armc-cc6 |
72 | 786.6 MiB |
runs/sft_first_decl-cc |
77 | 38.2 MiB |
runs/sft_first_decl-cc2 |
58 | 24.5 MiB |
runs/sft_first_decl-cc3 |
8 | 3.5 MiB |
runs/sft_first_decl-cc4 |
45 | 211.3 MiB |
runs/sft_first_idd |
7 | 1.9 MiB |
runs/sft_first_idd-cc |
97 | 1.8 GiB |
runs/sft_freq_decl-cc |
8 | 5.1 MiB |
runs/sft_inter_decl-cc |
8 | 5.1 MiB |
runs/sft_inter_decl-cc2 |
41 | 18.4 MiB |
runs/sft_interleaved |
24 | 5.1 MiB |
runs/sft_interleaved-claimform |
41 | 19.1 MiB |
runs/sft_joint |
24 | 5.0 MiB |
runs/sft_joint-cc |
47 | 17.2 MiB |
runs/sft_joint-claimform |
53 | 38.3 MiB |
runs/sft_joint-v2 |
19 | 3.4 MiB |
runs/shakedown |
192 | 5.4 MiB |
runs/shakedown-baseline |
24 | 739.1 KiB |
runs/shakedown-sft_joint |
84 | 1.7 MiB |
slurm-logs |
6 | 9.9 MiB |
state |
70 | 444.0 KiB |
| total | 4014 | 8.9 GiB |
runs/<run>/
Loose: eval.yaml, shaping.json, play_train.jsonl / play_eval.jsonl (prompt rows),
qa_set.jsonl, recite_*.jsonl (recital probes), sft_data.jsonl where the run used chat rows,
eval_dumps/episodes.jsonl (one row per finished episode with its graded outcome) and
collapse_guard.json, meminfo/, debug_rollout/<id>.pt (the exact post-generation training
dump of each rollout: tokens, rollout log-probs, rewards — what --load-debug-rollout-data
replays), replay_rollout/ where a replay was used, qwen_md/.
Archived: qwen_work.tar.zst (one directory per episode: journal.jsonl from the MCP server,
transcript.jsonl, calls.jsonl, claude.log, settings.json, mcp.json) and
wandb_offline.tar.zst; wandb_export/<offline-run>/{history.jsonl,config.json,summary.json,run.json}
is the readable export of the latter (every logged training step and rollout metric).
| run family | runs | status |
|---|---|---|
| Claude-Code-harness arms (2026-09-14 … 19), the reported ones | baseline-cc (12 RL rollouts), sft_first_idd-cc, act_ctrl-cc, sft_first_armc-cc, sft_first_armc-cc6, sft_first_decl-cc4 |
valid; their checkpoints are the published models |
| declarative-row trials that collapsed | sft_first_decl-cc, sft_first_decl-cc2, sft_inter_decl-cc2 (-cc3, sft_inter_decl-cc, sft_freq_decl-cc: prepared, not played) |
valid as negative results |
| first harness arms with the 61-claim QA rows (2026-09-14) | sft_first-cc, sft_joint-cc, baseline-cc-old-20260914 |
superseded (rows later audited) |
| in-process MCP client arms (2026-09-10 … 13) | baseline, sft_first, sft_interleaved, sft_joint, *-claimform, *-v2, sft_first-v3, baseline-shaped, sft_first-curatedL1 |
superseded by the harness; the claimform arms collapsed |
| teacher harvests that built the harness rows | cd82-play-v2, -v3, -v4, -v4a, -v4b (sft/rows*.jsonl.gz, build_report, coverage.json, SHA256SUMS) |
data builds; cd82-play-v4/sft = the IDD / act-ctrl / paired rows, cd82-play-v4/sft_armc = the arm-C rows, sft_nostance = the stance-stripped pair |
| shakedowns, probes, failed starts | shakedown*, dbg-*, baseline-qc*, baseline-memrepro-*, memrepro, mask_knowledge, curatedL1_check, *-fail-*, cd82-play-v1-fail-* |
diagnostics only |
metrics/
slurm_rl_metrics.jsonl — every rollout N: {…}, perf N: {…} (play metrics e2k/*) and
eval N: {…} dictionary printed by the jobs, keyed by job id and run; slurm_rl_metrics.md —
the per-run tables of clear rate, shaped return, step change / no-op fractions, think length and
recital fractions per rollout.
checkpoints/
checkpoints.md / checkpoints.json — all 84 Megatron torch_dist iteration directories
(27.5 TiB; 351 GB each with optimizer state) with file counts, sizes, completeness and the
run's latest_checkpointed_iteration. They stay on Isambard-AI. Every complete iteration of
every run (77 directories) was converted to bf16 safetensors and published as its own model
repository, ziyan98/E2K-cd82-Qwen3.8-27B-<run>-iter<N> (-frozen-adam suffix for the first,
frozen-Adam attempt of the IDD and act-ctrl runs); series.md lists them with the phase of each
iteration (SFT-only warm-up vs RL). The SFT-only iterations form, per arm, the series "more
supervised knowledge, same base": arm C sft_first_armc-cc 0/1/2 then sft_first_armc-cc6 3/4/5
(cite weight 1 → 8 → 32 → 32 → 32 → 32, knowledge NLL 3.05 → 0.42 per token), IDD 0/1, act-ctrl
0/1, declarative trials 0–2 or 0–3. Each repository also carries megatron/common.pt and
args.json (every Megatron/slime argument of that checkpoint), latest_checkpointed_iteration.txt
and the rollout data cursor, so a run can be rebuilt with slime's convert_hf_to_torch_dist.py
(optimizer moments are not published). The reported policies:
| model repository | run / iteration | what the policy is |
|---|---|---|
E2K-cd82-Qwen3.8-27B-sft_first_armc-cc6-iter6 |
arm C, 6 SFT warm-ups + 1 RL rollout | knowledge NLL 0.42, first verbatim recital in play |
E2K-cd82-Qwen3.8-27B-sft_first_armc-cc6-iter5 |
arm C, SFT-only | the policy that played RL rollout 6 |
E2K-cd82-Qwen3.8-27B-sft_first_idd-cc-iter8 |
IDD, 2 SFT + 7 RL rollouts | best RL trajectory (+0.25) |
E2K-cd82-Qwen3.8-27B-sft_first_idd-cc-iter1 |
IDD, SFT-only | |
E2K-cd82-Qwen3.8-27B-baseline-cc-iter11 |
pure RL, 12 rollouts | |
E2K-cd82-Qwen3.8-27B-act_ctrl-cc-iter3 |
act-ctrl control, 2 SFT + 2 RL | |
E2K-cd82-Qwen3.8-27B-act_ctrl-cc-iter1 |
act-ctrl, SFT-only | |
E2K-cd82-Qwen3.8-27B-sft_first_decl-cc4-iter2 |
declarative rows (tool-call ending), SFT-only | |
E2K-cd82-Qwen3.8-27B-sft_first_decl-cc4-iter3 |
same + 1 RL rollout | |
E2K-cd82-Qwen3.8-27B-sft_first_armc-cc-iter2 |
arm C, 3 SFT warm-ups (knowledge NLL 1.56) |
Other directories
knowledge-set/20260912-0920-final/— the atoms, reports, recall forms andrl_config_*.jsonthe arms were launched with (the full build with ledgers is in the harness dataset).state/— the run logearly_recital_snapshot_20260916.md, job-id files, the arm-C row build scripts, the source-rule trial reports, relaunch scripts.numeric/— the numerical-correctness hunt that preceded the runs: TransformerEngine's cuDNN fused attention has a wrong backward for packedthd+ head_dim 256 (flash-attn 2.8.3 is correct), fla's gated-delta-rule backward under triton 3.7.1 is clean, SGLang's Qwen3.5 forward is wrong under tensor parallelism > 1 on this image (tp=1 matches HF); scripts, logs and the agent-diagnostic transcripts (dbg/), without the multi-GB tensor dumps.slurm-logs/— RL job logs (186 files), diagnostics, data-build and conversion job logs.
Caveats
- One game, one model, 32 episodes per rollout: every effect is on a single seed band and the rollout-to-rollout noise is about 0.05 shaped return.
- The compute award ran out on 2026-09-18/19:
sft_first_armc-cc6stops after one RL rollout andsft_first_idd-ccafter seven; the prepared continuation configs (rl_config_cc_armc_inter.json,rl_config_cc_armc_freq.json,rl_config_cc_idd_v2.json) were never run.
Integrity and licence
MANIFEST.sha256 / MANIFEST.json hash every file. CC BY 4.0. Ziyan Wang, King's College
London, 2026; compute: Isambard-AI (BriCS, University of Bristol).
- Downloads last month
- 121