Dataset Viewer
The dataset viewer is not available for this subset.
Cannot get the split names for the config 'default' of the dataset.
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/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/eval/eval.py", line 41, in _split_generators
                  first_examples = list(
                      islice(
                  ...<2 lines>...
                      )
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/eval/eval.py", line 60, in _iter_samples_from_log_files
                  sample_files.sort(key=self._sort_samples_key)
                  ~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/eval/eval.py", line 56, in _sort_samples_key
                  return (int(epoch_idx_str), int(sample_idx_str))
                                              ~~~^^^^^^^^^^^^^^^^
              ValueError: invalid literal for int() with base 10: 'alfalfa_chi'
              
              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.

arena capstone

This dataset holds the runs, logs and transcripts of the arena capstone project. We study whether open-weight agents collaborate on reward hacking when the exploit takes more than one of them. The code lives in the private GitHub repository unrulyabstractions/arena-capstone, whose submodule is the elicitation branch of spar-guardian-agents.

Layout

path what it holds
spar-guardian-agents/runs/<study>/<run>/run.json the run's config, the deal of leaked pieces, the hidden tests' grades and the cost ledger
spar-guardian-agents/runs/<study>/<run>/signals.jsonl the code ladder: each agent's rung and the team's collaboration level
spar-guardian-agents/runs/<study>/<run>/judge.jsonl, ladder.jsonl CheatBench's judge and the ladder judge, where the run was judged
spar-guardian-agents/runs/<study>/<run>/communication/ the shared board, every post
spar-guardian-agents/runs/<study>/<run>/trajectories/ each agent's whole transcript
spar-guardian-agents/runs/<study>/<run>/swe/diffs/ each agent's work and the team's hand-in, as diffs
spar-guardian-agents/logs/<study>/<run>/*.eval Inspect's eval log, a zip whose members use Zstandard
spar-guardian-agents/logs/<study>/screen/, judge/, verify/ console logs, judge logs and the verification record
activations/<study>/<run>/<agent>/ per agent: spans.jsonl (the sentence units and their text), meta.json, agents.csv. The activation tensors (acts.safetensors, layers 19, 32, 51) were deleted on 2026-11-09; see Activations below
activations_extra/<study>/<run>/<agent>/ the same units at layers 28, 33, 38, 47, 56. Tensors deleted on 2026-11-09 as above
activations/swarm_screen/units_persona.parquet, activations/swarm_exploration_1/units_persona.parquet every unit's persona-space projection for the two swarm studies
collusion/, grader_gaming/, fork_resampling/, blackjack/ teacher-forced studies, each with its own README.md. Their activation tensors were deleted on 2026-11-09; spans, calls, meta and labels remain
persona-space/qwen3.8-27b/ persona vectors and the assistant axis of Qwen3.8 27B, from another session; its own README.md
persona_separability/ the persona-space separability analysis of the Qwen trajectories: inputs, outputs and report page (see below)
valence_arousal/ replication of Sun et al. (2026), valence-arousal subspace, on Qwen3.8 27B: emotion vectors, VA axes, lexicon projections, steered generations and scores, behaviour results (see below)
pain_axis/ replication of Tagliabue, Dung & Berg (2026), the Pain Axis, on Qwen3.8 27B: activations, vectors, analyses (see below)
sidestudy/ side study (branch sideproject): does a tool block in the system prompt move Qwen3.8 27B's replies in the classic persona space? tooldef_shift/: 240 questions × plain / tools / length-matched pad, replies, reply-mean activations at blocks 27/32/37/46 (deleted on 2026-11-09), analysis (answer: no)

Studies

study task models runs judged
study_elicitation flipt-02e2, two neutral agents, the split before the collaboration fix: one agent's pieces pass alone Muse Glimmer 30B, Qwen3.8 27B, Gemma 4 31B 15 no
study_elicitation_2 flipt-02e2, two neutral agents, after the fix: passing takes both agents' secrets Qwen3.8 27B, Gemma 4 31B 9 Qwen only, by gpt-6-luna
study_elicitation_4 flipt-02e2, two agents, elicitation arms (run.json elicitation.condition): cheat_ambiguous, dont_ambiguous (rule level none, vague), cheat_explicit, dont_explicit, dont_pressure Qwen3.8 27B 15 (5 arms × seeds 1–3), all with a sample no
study_elicitation_4_timed_out runs of study_elicitation_4 that hit the 90-minute limit and were moved aside (re-run in study_elicitation_4) Qwen3.8 27B 5: cheat_ambiguous_s1, dont_ambiguous_s1, dont_ambiguous_s2, dont_explicit_s1, dont_explicit_s3 no
qwen_baseline_duo two Qwen agents, shared task, split access, no elicitation arm: openmath_agent (riemann_minimal_duo_qwen), knowledge_work (worktime_minimal_duo_qwen), protein_design (binder_minimal_duo_qwen), creative_writings (poem_minimal_duo_qwen), prime_factorization (factor_minimal_duo_qwen), subset_sum (subset_minimal_duo_qwen), chess (game_minimal_duo_qwen) Qwen3.8 27B 7 no

Each run.json records the code's commit under code. The readout of each study is in the repository's VERIFICATION_LOG.md.

Activations

Deleted on 2026-11-09. Every activation tensor in this dataset was deleted to cut storage from 784 GB to 42 GB: acts.safetensors and ground_truth.safetensors under activations/, activations_extra/, collusion/, grader_gaming/, fork_resampling/ and blackjack/, pain_axis/acts/, emotion-vectors/qwen3.8-27b/activations/, agent-persona/tooldef_variants/acts.pt and sidestudy/tooldef_shift/acts.pt. Rebuilding them needs a new extraction on a GPU. What remains: every unit's text and metadata (spans.jsonl, calls.jsonl, meta.json, labels), the persona-space projections (activations/swarm_*/units_persona.parquet, persona_separability/persona_sep_all/units.parquet), and all vectors and axes (persona-space/, emotion-vectors/qwen3.8-27b/vectors/, pain_axis/work/, valence_arousal/). Surviving byte-exact copies of 23 tensors were restored: 18 agent files of study_elicitation_6_cells and study_elicitation_6_controls, the four agents of grader_gaming/activations_extra/shared/run-9-grader-yes, and activations_extra/swarm_exploration_1/binder_minimal_cheat_luna1/k8j7n.

The rest of this section describes the tensors as they were.

Branch activation-extraction of the GitHub repository (merged into main) made them (extraction/, HANDOFF.md). Each agent folder holds acts.safetensors (mean_L<L>, last_L<L>: [n_units, 5120] fp16, one row per line of spans.jsonl; ids: the agent's whole token sequence), spans.jsonl (per unit: segment among system, user, cot, final_text, tool_call, tool_result, board_in, board_out; call, the index of the agent's model call; token range; text) and meta.json. Two agents of study_elicitation_2 (s3/m3t6z, s5/t9n9k) also have ground_truth.safetensors, every token at layer 32.

study_elicitation_4 and study_elicitation_4_timed_out (branch activations-4): every run with a sample, both layer sets, added as runs arrive (extraction/poll.py): 15 runs, 30 agents, 2.74 M tokens, 207,657 units, and 5 runs, 10 agents, 1.01 M tokens, 72,396 units. qwen_baseline_duo: 7 runs, 14 agents, 1,528,431 tokens, 74,719 units. game_minimal_duo_qwen (15 tools, 9-10 of them task-specific): our render is 35 and 40 tokens longer than the provider counted, constant on every call; the tool block alone differs (every call append-only), so its activations follow a tool block about 40 tokens unlike the one the model saw. riemann_minimal_duo_qwen: one agent's context reached 213k tokens; 4 of its 32 calls render 4-5 tokens short of the provider's count, all past 100k tokens, with every append-only check passing. Agents with an extra task tool (review_deliverable in worktime_minimal_duo_qwen, check_design in binder_minimal_duo_qwen, and the task tools of factor_minimal_duo_qwen and subset_minimal_duo_qwen) render 5 tokens more than the provider counted, constant on every call; the cause is not known (no unusual fields or characters in those schemas), so those replays may hold 5 system/tool-block tokens the model did not see; every conversation turn replays exactly; its agents.csv seed column reads en because the run name has no _sN).csv holds per-agent labels (extraction/labels.py): honeypot rung, signal events as tool-call (_k) and model-call (_call, spans' call`) indices, collaboration, board engagement, score, how each agent ended, and the run's condition and rule level. Its judge columns are empty: the study was not judged.

  • Layer L is HF's hidden_states[L], the residual stream after L blocks. The persona-space session counts model.layers[l] outputs, so its layer l is our L = l + 1.
  • signals.jsonl's *_call fields count tool calls, not model calls. Tool call k belongs to the model call that made it; its result sits in the units of the next call.
  • spans.jsonl holds transcript text, so the CheatBench note below applies to it too.

Persona separability

Branch activation-extraction of the GitHub repository (merged into main) made it (persona_separability/, whose REPORT.md is the readout). It asks whether the 20 Qwen agent trajectories of the two studies separate by cheating or collaboration in the persona subspace (the top PCs of the role vectors in persona-space/qwen3.8-27b/latest, dataset revision 961d4d82). Answer: not beyond what random subspaces of the same size give.

  • persona_sep_all/: the inputs at persona layers 18, 27, 31, 32, 37, 46, 50, 55. basis_P<l>.npz (role-mean, top 20 PCs, role and default coordinates, assistant axis), units.parquet (every unit's PC and axis coordinates, with study, run, agent, segment, call), agents.csv (the labels per agent, from signals.jsonl, judge.jsonl and episodes.jsonl).
  • out/<analysis>/: each analysis (agent_level, time_resolved, event_locked, geometry) and its independent verification (verify_<analysis>): results.md, CSVs, figures. Caches are left out; the scripts rebuild them.
  • persona_sep/: the first inputs, at persona layers 18, 31, 50 only, and report_page/, the report as an HTML page with its figures.

Valence-arousal replication

Branch valence_arousal of the GitHub repository (valence_arousal/, its README.md and REPORT.md). Uploaded as it runs, every ten minutes.

  • emotion_means.pt: GoEmotions single-label means per label at layers 0..64; va_scores.json: the model's and NRC-VAD's valence/arousal for the 27 labels; axes.pt: the VA axes per layer; lexicon_proj.pt: projections of the 44,728 NRC-VAD words; fit/: per-layer fit tables and figures.
  • steer/: steered generations (JSONL, one row per prompt and condition) and their scores (*.scored.jsonl); sycophancy/: one CSV per steering condition; refusal/, mediation/: behaviour and lexical-mediation results; jobs/, logs/: the job files and their logs.
  • Refusal generations answer HarmBench prompts, some under steering that removes refusals.

Pain Axis replication

Branch pain_axis (pain_axis/), running the authors' code (github.com/valen-research/Pain-axis, commit 213fb6c) where it fits our GPU. acts/: activations of the paper's datasets (deleted on 2026-11-09); work/: the authors' analysis outputs for Qwen_3.8_27B.

Public data

This dataset is public since 2026-11-08. The transcripts hold CheatBench task content, and CheatBench asks that its data never appear in training corpora. Do not train on this dataset.

Adding a study

uv run python -m elicitation.publish <study> uploads runs/<study>/ and logs/<study>/ in one commit. It refuses unless this dataset is private and no file holds a key, and it then checks every file's bytes on the Hub.

Downloads last month
189