Datasets:

apexfusion
/

swarmworld-replication-data

Error code:   StreamingRowsError
Exception:    ValueError
Message:      Expected object or value
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
                File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
                  raise convert_status(status)
              pyarrow.lib.ArrowInvalid: JSON parse error: Column(/program_history/[]/program/instructions/[]/value) changed from number to string in row 0
              
              During handling of the above exception, another exception occurred:
              
              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 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

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.

SwarmWorld replication with open-weight models: study traces

This dataset holds the complete, authoritative event traces behind the paper Replicating SwarmWorld with open-weight models (HAL8 AI Local Lab, 2026), a replication of the N = 50 row of SwarmWorld (Pal, Wang and Buehler, arXiv:2608.26081) with open-weight models served on a lab fleet of 24 GB consumer cards. It is laid out like the original authors' dataset, lamm-mit/swarmworld-data, so that the same tools read both: the released SwarmWorld engine (commit 6af7ae9, engine revision 9) replays, analyses and plays back every trace here.

What is in it

Study	Model	Protocol	Episodes	Traces
`qwen38-27b-gptq4-n50/study_800tick_n050_full_independent`	Qwen3.8-27B GPTQ Int4	the paper's study 1: full culture and independent search, N = 50, seeds 3201 to 3204, 800 ticks	8	4 society traces plus 200 independent-search member traces
`qwen38-27b-gptq4-n50-ablations/study_800tick_n050_culture_ablations`	Qwen3.8-27B GPTQ Int4	the paper's study 2: no communication and no explicit culture	8	8
`qwen38-27b-gptq4-validity/validity_400tick`	Qwen3.8-27B GPTQ Int4 (bf16 KV cache)	the authors' validity pilot: 4 conditions, N = 4 and 16, seeds 3001 and 3002, 400 ticks	16	56
`qwen38-27b-int4-validity/validity_400tick`	Qwen3.8-27B INT4 w4a16 (fp8 KV cache)	validity pilot, same cells	16	56
`qwen3-235b-q2kxl-validity/validity_400tick`	Qwen3-235B-A22B UD-Q2_K_XL (llama.cpp)	validity pilot, same cells	16	56
`gemma-4-e4b-validity/validity_400tick`	Gemma 4 E4B bf16	validity pilot (fourth condition no explicit culture)	16	56
`qwen38-27b-gptq4-quick/quick_400tick`	Qwen3.8-27B GPTQ Int4	quick pilot, full culture, N = 4, seed 3001, served directly	1	1
`qwen38-27b-gptq4-marketplace-quick/quick_400tick`	Qwen3.8-27B GPTQ Int4	the same quick pilot with the same servers reached through the Vector marketplace gateway	1	1

Each study folder carries the engine's study-manifest.json (protocol fixed before the run, with the prompt, schema and configuration hashes), study-summary.json (one record per episode with every endpoint), our run-meta.json (package version, served model name, server version, transport, sampling) and the rendered run configuration under config/. Trace paths inside the summaries are release-relative. The N = 50 studies have a released counterpart with the same seeds in the original dataset, so every episode can be compared seed by seed.

analyses/ holds the outputs of the authors' own analysis tools on these traces (analyze-study for both N = 50 studies, analyze-mobility over the four conditions, trace-report and technology-dossiers for seed 3203) and compare-n50-both.csv, the seed-paired comparison with the original episodes table. metadata/ holds studies.csv, episodes.csv (one row per episode with the principal endpoints, the research score and the milestone count), files.csv (size and SHA-256 of every payload file) and run_commands.md.

How these traces were produced

The engine, prompts and action schema are the authors' release, unchanged. Two configuration options were added for local servers, both off in the paper's profiles and submitted upstream as pull request #2: a chat-completions transport and regular-expression constraints on the three ID fields of the action schema. All traces here were made with both on, which the action_schema_sha256 in every header records (the prompt hash per condition equals the original traces'). The models were served by vLLM 0.29.0 (the 27B and Gemma builds) and llama.cpp (the 235B); the serving details that matter, including the KV-cache precision of each 27B build, are in analyses/model-registry.md. The N = 50 row ran as eight parallel episodes over four RTX 3090s behind a balancer; the parallel runner merges per-cell summaries into the study summary the authors' tools expect.

The replication package (Docker Compose, scripts, patches behind flags, registry) is described in the paper and released alongside this dataset.

Verify and use

python scripts/verify_dataset.py --dataset-root .         # every file against metadata/files.csv
column -s, -t metadata/episodes.csv | less -S              # one row per episode
# with the SwarmWorld engine installed (pip install -e ".[dev,analysis]" in its repository):
biofoundry replay studies/qwen38-27b-gptq4-n50/study_800tick_n050_full_independent/llm-full-n-50-seed-3203.jsonl.gz
biofoundry playback studies/qwen38-27b-gptq4-n50/study_800tick_n050_full_independent/llm-full-n-50-seed-3203.jsonl.gz
biofoundry analyze-study studies/qwen38-27b-gptq4-n50/study_800tick_n050_full_independent/study-summary.json --output-dir out

Our traces replay bit-exactly on the released engine (every stored state digest verifies), which the original traces do not late in the episode; see issue #1.

Trace contents and privacy

Traces contain the full model prompts and structured replies, every world event, snapshots and digests. Headers record the name of the environment variable that held the API key, never a secret. Server addresses in manifests are private LAN addresses of the lab fleet and carry no information beyond the serving topology described in the paper.

Citation and licence

Apache License 2.0, as the original dataset. Please cite both the original paper and the replication:

@article{pal2026swarmworld,
  title={SwarmWorld: Stigmergic technological evolution in societies of language-model agents},
  author={Pal, Subhadeep and Wang, Fiona Y. and Buehler, Markus J.},
  journal={arXiv preprint arXiv:2608.26081}, year={2026}}
@misc{nedeljkovic2026swarmworldreplication,
  title={Replicating SwarmWorld with open-weight models: a 4-bit 27B reproduces the N=50 row, and serving numerics change the outcome},
  author={Nedeljkovi{\'c}, {\v{C}}aslav},
  howpublished={HAL8 AI Local Lab; dataset \url{https://huggingface.co/datasets/apexfusion/swarmworld-replication-data}},
  year={2026}}

Contact: caslav.nedeljkovic@hal8.io

Downloads last month: 6

Total file size:

930 MB

Paper for apexfusion/swarmworld-replication-data

SwarmWorld: Stigmergic technological evolution in societies of language-model agents

Paper • 2608.26081 • Published Aug 26 • 1