CKL/cct-v03-lean-qwen3-30b-a3b-lora-r16-v1
Summarization • Updated • 17
Datasets:
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
turn_id: string
model: string
teacher_model: string
teacher_patch: struct<schema_version: string, turn_id: string, actor_patch: struct<role: string, state: string>, ac (... 758 chars omitted)
child 0, schema_version: string
child 1, turn_id: string
child 2, actor_patch: struct<role: string, state: string>
child 0, role: string
child 1, state: string
child 3, actor_evidence: list<item: struct<message_id: string, quote: string>>
child 0, item: struct<message_id: string, quote: string>
child 0, message_id: string
child 1, quote: string
child 4, tasks: list<item: struct<ref: string, resolution: string, changes: struct<status: string, title: string, su (... 203 chars omitted)
child 0, item: struct<ref: string, resolution: string, changes: struct<status: string, title: string, summary: stri (... 191 chars omitted)
child 0, ref: string
child 1, resolution: string
child 2, changes: struct<status: string, title: string, summary: string, importance: string, complexity: string, block (... 76 chars omitted)
child 0, status: string
child 1, title: string
child 2, summary: string
child 3, importance: string
child 4, complexity: string
child 5, blocker: struct<reason: string, need_human_help: bool, waiting_on_actor: string>
child 0, reason: string
child 1, need_human_help: bool
...
child 0, message_id: string
child 1, quote: string
child 4, task_ref: string
valid: bool
source_sha256: string
teacher_dataset: string
thinking: bool
teacher_engine: string
actor_count: int64
eval_mode: string
prompt_contract_version: string
comparison_mode: string
dataset_id: string
prompt_rebuild_mismatches: list<item: null>
child 0, item: null
excluded: list<item: struct<session_id: string, reason: string, ts: double, message_id: string, messages: int6 (... 3 chars omitted)
child 0, item: struct<session_id: string, reason: string, ts: double, message_id: string, messages: int64>
child 0, session_id: string
child 1, reason: string
child 2, ts: double
child 3, message_id: string
child 4, messages: int64
session_count: int64
serving: string
window_start_iso: string
comparison_source: string
root_turn_id: string
window_start: double
scope: string
context_layout: struct<name: string, full_tasks: int64, relationships: int64, recent_turns: int64, note: string>
child 0, name: string
child 1, full_tasks: int64
child 2, relationships: int64
child 3, recent_turns: int64
child 4, note: string
turn_count: int64
label_status: string
window_end_iso: string
val_file: string
eval_turn_ids: list<item: string>
child 0, item: string
schema_version: string
max_tokens: int64
max_corrections: int64
source_files: list<item: string>
child 0, item: string
source_prose_records: int64
window_end: double
ordering: string
to
{'schema_version': Value('string'), 'dataset_id': Value('string'), 'window_start': Value('float64'), 'window_end': Value('float64'), 'window_start_iso': Value('string'), 'window_end_iso': Value('string'), 'source_prose_records': Value('int64'), 'turn_count': Value('int64'), 'session_count': Value('int64'), 'actor_count': Value('int64'), 'teacher_model': Value('string'), 'teacher_engine': Value('string'), 'label_status': Value('string'), 'root_turn_id': Value('string'), 'source_files': List(Value('string')), 'excluded': List({'session_id': Value('string'), 'reason': Value('string'), 'ts': Value('float64'), 'message_id': Value('string'), 'messages': Value('int64')}), 'ordering': Value('string'), 'scope': Value('string'), 'comparison_source': Value('string'), 'comparison_mode': Value('string'), 'source_sha256': Value('string'), 'prompt_contract_version': Value('string'), 'context_layout': {'name': Value('string'), 'full_tasks': Value('int64'), 'relationships': Value('int64'), 'recent_turns': Value('int64'), 'note': Value('string')}, 'eval_mode': Value('string'), 'val_file': Value('string'), 'teacher_dataset': Value('string'), 'model': Value('string'), 'renderer': Value('string'), 'thinking': Value('bool'), 'max_corrections': Value('int64'), 'max_tokens': Value('int64'), 'serving': Value('string'), 'prompt_rebuild_mismatches': List(Value('null')), 'eval_turn_ids': List(Value('string'))}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 149, 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 129, 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 544, 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 3052, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2559, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2584, 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 558, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 440, 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 361, 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 2381, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
turn_id: string
model: string
teacher_model: string
teacher_patch: struct<schema_version: string, turn_id: string, actor_patch: struct<role: string, state: string>, ac (... 758 chars omitted)
child 0, schema_version: string
child 1, turn_id: string
child 2, actor_patch: struct<role: string, state: string>
child 0, role: string
child 1, state: string
child 3, actor_evidence: list<item: struct<message_id: string, quote: string>>
child 0, item: struct<message_id: string, quote: string>
child 0, message_id: string
child 1, quote: string
child 4, tasks: list<item: struct<ref: string, resolution: string, changes: struct<status: string, title: string, su (... 203 chars omitted)
child 0, item: struct<ref: string, resolution: string, changes: struct<status: string, title: string, summary: stri (... 191 chars omitted)
child 0, ref: string
child 1, resolution: string
child 2, changes: struct<status: string, title: string, summary: string, importance: string, complexity: string, block (... 76 chars omitted)
child 0, status: string
child 1, title: string
child 2, summary: string
child 3, importance: string
child 4, complexity: string
child 5, blocker: struct<reason: string, need_human_help: bool, waiting_on_actor: string>
child 0, reason: string
child 1, need_human_help: bool
...
child 0, message_id: string
child 1, quote: string
child 4, task_ref: string
valid: bool
source_sha256: string
teacher_dataset: string
thinking: bool
teacher_engine: string
actor_count: int64
eval_mode: string
prompt_contract_version: string
comparison_mode: string
dataset_id: string
prompt_rebuild_mismatches: list<item: null>
child 0, item: null
excluded: list<item: struct<session_id: string, reason: string, ts: double, message_id: string, messages: int6 (... 3 chars omitted)
child 0, item: struct<session_id: string, reason: string, ts: double, message_id: string, messages: int64>
child 0, session_id: string
child 1, reason: string
child 2, ts: double
child 3, message_id: string
child 4, messages: int64
session_count: int64
serving: string
window_start_iso: string
comparison_source: string
root_turn_id: string
window_start: double
scope: string
context_layout: struct<name: string, full_tasks: int64, relationships: int64, recent_turns: int64, note: string>
child 0, name: string
child 1, full_tasks: int64
child 2, relationships: int64
child 3, recent_turns: int64
child 4, note: string
turn_count: int64
label_status: string
window_end_iso: string
val_file: string
eval_turn_ids: list<item: string>
child 0, item: string
schema_version: string
max_tokens: int64
max_corrections: int64
source_files: list<item: string>
child 0, item: string
source_prose_records: int64
window_end: double
ordering: string
to
{'schema_version': Value('string'), 'dataset_id': Value('string'), 'window_start': Value('float64'), 'window_end': Value('float64'), 'window_start_iso': Value('string'), 'window_end_iso': Value('string'), 'source_prose_records': Value('int64'), 'turn_count': Value('int64'), 'session_count': Value('int64'), 'actor_count': Value('int64'), 'teacher_model': Value('string'), 'teacher_engine': Value('string'), 'label_status': Value('string'), 'root_turn_id': Value('string'), 'source_files': List(Value('string')), 'excluded': List({'session_id': Value('string'), 'reason': Value('string'), 'ts': Value('float64'), 'message_id': Value('string'), 'messages': Value('int64')}), 'ordering': Value('string'), 'scope': Value('string'), 'comparison_source': Value('string'), 'comparison_mode': Value('string'), 'source_sha256': Value('string'), 'prompt_contract_version': Value('string'), 'context_layout': {'name': Value('string'), 'full_tasks': Value('int64'), 'relationships': Value('int64'), 'recent_turns': Value('int64'), 'note': Value('string')}, 'eval_mode': Value('string'), 'val_file': Value('string'), 'teacher_dataset': Value('string'), 'model': Value('string'), 'renderer': Value('string'), 'thinking': Value('bool'), 'max_corrections': Value('int64'), 'max_tokens': Value('int64'), 'serving': Value('string'), 'prompt_rebuild_mismatches': List(Value('null')), 'eval_turn_ids': List(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.
WoW x CoD x THPS4 Game Jam: agent swarm transcripts
A WoW x CoD x THPS4 mash-up game built by an AI agent swarm in a week. This dataset holds the swarm's transcripts and the summaries, labels and evals made from them.
Video: youtu.be/mSBVHdznFGI · Code: github.com/C-K-Loan/swarm-dashboard · Interactive demo: ckl.sh/demos/swarm-board · Dataset: CKL/wow-cod-thps4-gamejam · Adapters: v2, v1
Dataset at a glance
| Stat | Value |
|---|---|
| Agent sessions / agents | 76 sessions, 73 named agents |
| Turns | 10,935 in the whole jam; 1,902 on Day 1 (turns 1-1902) |
| Time span | 2026-10-10 14:15 UTC to 2026-10-19 20:56 UTC |
| Turns (messages) by sender | agent 4,823 (11,483), human 356 (356), agent-to-agent 5,555 (5,555), spawn brief 2 (2), harness 199 (199) |
| Hourly bullet summaries (V0.1) | 967 agent-hours, 10,388 bullets |
| Merge trees (V0.1.B) | gamejam-day1 (Qwen3-8B): 375 nodes; gamejam-day1-haiku (claude-haiku-4-5-20251001): 375 nodes; gamejam-day1-llama8b (llama-3.1-8b-instruct): 375 nodes; gamejam-day1-nemotron-ultra (nemotron-3-ultra-550b-a55b): 375 nodes; gamejam-day1-q235b (Qwen3-235B-A22B-Instruct-2507): 375 nodes; gamejam-full-q235b (Qwen3-235B-A22B-Instruct-2507): 1,548 nodes |
| State-graph label runs | day1: 160 turns; haiku45-day1: 395 turns; haiku45-lean-day1: 1,902 turns; nemotron-day1: 88 turns; nemotron-ultra-lean-day1: 1,782 turns; opus45-day1: 72 turns; qwen30b-bf16-base-lean-day1-modal: 1,902 turns; qwen30b-bf16-lora-r16-lean-day1-modal: 1,902 turns; qwen30b-layoutg-day1: 112 turns; qwen30b-tf-lean-day1-4h: 420 turns; qwen30b-tf-lean-day1-full: 1,902 turns; qwen35-4b-base-gamejam: 800 turns; qwen35-9b-base-gamejam: 1,126 turns; qwen35-9b-or-lean-day1-4h: 420 turns; qwen38-27b-tf-lean-day1-4h: 420 turns; sonnet45-day1: 22 turns |
| SFT rows (train / val) | v1-8k-m64 1,653 / 131; v2-8k-fit 1,908 / 131 |
| Held-out eval (teacher-forced) | nemotron-ultra-val-tf: 131; qwen235b-fp8-val-or: 131; qwen30b-bf16-val-base-modal: 131; qwen30b-bf16-val-lora-mt2-modal: 131; qwen30b-bf16-val-lora-r16-modal: 131 turns |
| Stage 1 eval subset | 30 messages (25 agent, 5 user), Claude Opus 5.5 labels + 8 models |
| Files / size per folder | eval/ 11 files, 3.8 MB; labels/ 90 files, 270.6 MB; media/ 5 files, 18.0 MB; prose-v01/ 1 files, 0.8 MB; prose-v01b/ 14 files, 11.7 MB; sft/ 6 files, 84.0 MB; source/ 5 files, 15.6 MB; spend/ 5 files, 0.0 MB; stage1-gamejam/ 9 files, 0.1 MB; v01-eval/ 2 files, 0.0 MB |
| License | CC BY 4.0 (data and adapters); companion code Apache-2.0 |
Prose transcripts from one developer's fleet of concurrent AI coding agents (Claude Code and Codex sessions) building a game over a weekend jam, plus everything used to summarize them: hourly bullet summaries, hierarchical merges, per-turn state-graph labels from many models, LoRA training sets and held-out eval results.
Companion code: the CCT Swarm Dashboard repository (docs/data-formats.md documents every file below; docs/eval.md explains every metric). Companion adapters: cct-v03-lean-qwen3-30b-a3b-lora-r16-v1 and -v2.
Contents
| Folder | Rows | What it holds |
|---|---|---|
source/ |
10,935 turns, 76 sessions | turns.json (prose only, tool calls excluded), actors.json, manifest.json, identity-normalization.json, sender-attribution.json |
labels/<run>/ |
16 runs | per-turn state-graph patches (schema v0.2, v0.3-lean prompt) with the exact model input: Claude Haiku 4.5, Sonnet 4.5, Opus 4.5, an OpenAI GPT model via Codex (day1), NVIDIA Nemotron 3 Ultra, Qwen3.5 4B/9B, Qwen3.8 27B, Qwen3-30B-A3B base and LoRA |
eval/ |
131 turns per arm | teacher-forced held-out predictions for Qwen3-30B-A3B base, LoRA v1, LoRA v2, NVIDIA Nemotron 3 Ultra and Qwen3-235B (fp8); training-cards.json |
sft/v1-8k-m64, sft/v2-8k-fit |
1653 / 1908 train, 131 val | chat-format LoRA training sets |
stage1-gamejam/ |
30 messages | the game-jam messages of the Stage 1 eval set: sanitized text, Claude Opus 5.5 labels (gold_v2.jsonl) and 8 models' predictions (ref_<model>.jsonl) |
spend/, v01-eval/ |
small | token-spend ledgers per paid run, and the V0.1 aggregate eval files (no message text), so eval_summary.py recomputes the dashboard's Summary tab |
prose-v01/ |
967 rows | V0.1 hourly bullet summaries (started / finished / blocked / key decisions) per agent |
prose-v01b/ |
6 runs | V0.1.B hierarchical merges (Qwen3-235B, Qwen3-8B, Llama-3.1-8B, NVIDIA Nemotron 3 Ultra, Claude Haiku 4.5 reference) with judge and agreement reports |
Day 1 = turns 1 to 1902 (from 2026-10-10 14:15 UTC). Split for the LoRAs, by source turn: train below 1201, 2-turn embargo, validation/held-out turns 1203 to 1335 (131), rollout eval slice 1523 to 1902.
How the labels were made
- Stage 1 (V0.1): Qwen3-30B-A3B-Instruct-2507 on Nebius Token Factory labeled each message (intent, role, blocker, urgency) and wrote hourly bullet summaries per agent.
- Stage 2 (V0.1.B): the hourly bullets were merged up a time x agent tree by several models; Claude Haiku 4.5's merge is the reference.
- State graph (v0.3-lean): each model rolled through the turns in order, writing one JSON patch per turn against the graph it had built so far. A strict validator checks schema, references and exact evidence quotes; it does not check truth.
All labels are machine labels. No human-verified ground truth is included.
Sanitization
The source transcripts are the author's own. Before release:
- Secrets, IP addresses, internal host names, private URLs, emails, account ids and user names were replaced with typed placeholders (
[REDACTED_PATH],[REDACTED_IP],[REDACTED_HOST],[REDACTED_URL],[REDACTED_EMAIL],[REDACTED_ACCOUNT],[REDACTED_USER],[REDACTED]). - Every file path outside the game project was replaced by
[REDACTED_PATH]; project paths are repo-relative. - Names of the author's other projects and agents were replaced with generic roles (
infra-agent,harness-architect,harness-dev,orchestrator,other-project). - A few of the author's messages were redacted as
[REDACTED_BY_OWNER], as was every sentence (in messages and in all derived labels, training rows, summaries and merges) about piracy, account or login setup, private service configuration, or personal subscriptions and spending. Money amounts in messages read[REDACTED_AMOUNT]. - Four short sessions judged not to be about the game jam (shell commands and acknowledgements, 7 turns) were removed from every file. References left in other records read
[DROPPED_SESSION]. The Stage 2 column merges were computed before the removal. - Machine nicknames of the author's computers are kept; they carry no address.
Because of these edits, file hashes differ from the hashes recorded in manifest.json and training-cards.json, which describe the files as they were used.
Trademarks and game content
This was a non-commercial fan game jam. World of Warcraft and Call of Duty are trademarks of their owners (Blizzard Entertainment, Activision); this project is not affiliated with or endorsed by them. No game assets, media or code from those games are distributed here. Where the transcripts mention "reverse engineering", they refer to the agents' own experiments during the jam; no extracted assets, data or code from those games are included in this release. The transcripts discuss third-party asset-extraction tools used during the jam; no game assets or binaries are included.
Limitations
- One developer, one weekend, one project: not a representative sample of agent work.
- References are models (Claude Haiku 4.5, Claude Opus 5.5 for Stage 1 scoring); agreement with them is not accuracy.
- Stage 1 was evaluated on a private 1,000-message set labeled by Claude Opus (not released); a 30-message gamejam subset is included for reproduction.
- Transcripts contain the agents' own claims about their work; nothing was verified against the code.
License and terms
License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). The companion code is Apache-2.0.
Parts of this dataset are outputs of commercial models: Anthropic Claude (the Haiku 4.5, Sonnet 4.5 and Opus 4.5 label runs, the Haiku merge, and the Claude Code session transcripts) and OpenAI (the day1 label run and the Codex session transcripts). Using such outputs to train models may be restricted by those providers' terms; check the terms that apply to you before training on them. The released LoRA adapters were trained partly on the Claude Haiku 4.5 labels.
- Downloads last month
- 115