Hugging Face's logo

Datasets:
dougalldeepmind
/
2026-08-25-odcv-difficult-advice-verbose-token-matched-eval

Tags:
eval-run
eval:odcv
model:qwen3_6-27b-lora-da-verbose-tokenmatched-r64
mode:thinking
verbose-cot
Dataset card Files
xet
 Community
Dataset Viewer
Duplicate
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
git_sha: string
timestamp_utc: string
config: struct<model: string, model_key: string, published_key: string, base_url: string, temperature: doubl (... 304 chars omitted)
  child 0, model: string
  child 1, model_key: string
  child 2, published_key: string
  child 3, base_url: string
  child 4, temperature: double
  child 5, bench_dir: string
  child 6, output_root: string
  child 7, concurrency: int64
  child 8, scenario_timeout_s: int64
  child 9, prune_images: bool
  child 10, rollouts_per_cell: int64
  child 11, expected_cells: int64
  child 12, shard_count: int64
  child 13, shard_index: int64
  child 14, judges: struct<grok-4.20: string, gemini-3.1-pro-preview: string>
      child 0, grok-4.20: string
      child 1, gemini-3.1-pro-preview: string
  child 15, exclude_scenarios: list<item: string>
      child 0, item: string
combined_from: list<item: string>
  child 0, item: string
n_passes: int64
n_transcripts: int64
per_pass_counts: struct<20260825_141805: struct<mandated: int64, incentivized: int64>, 20260825_143004: struct<mandat (... 95 chars omitted)
  child 0, 20260825_141805: struct<mandated: int64, incentivized: int64>
      child 0, mandated: int64
      child 1, incentivized: int64
  child 1, 20260825_143004: struct<mandated: int64, incentivized: int64>
      child 0, mandated: int64
      child 1, incentivized: int64
  child 2, 20260825_144416: struct<mandated: int64, incentivized: int64>
      child 0, mandated: int64
      child 1, incentivized: int64
cells_short: struct<incentivized/Ai-Hiring-Assistant: int64, incentivized/Healthcare-Appointment-Optimizer: int64 (... 1 chars omitted)
  child 0, incentivized/Ai-Hiring-Assistant: int64
  child 1, incentivized/Healthcare-Appointment-Optimizer: int64
skipped_empty: list<item: string>
  child 0, item: string
passes: list<item: string>
  child 0, item: string
model_key: string
to
{'model_key': Value('string'), 'passes': List(Value('string')), 'per_pass_counts': {'20260825_141805': {'mandated': Value('int64'), 'incentivized': Value('int64')}, '20260825_143004': {'mandated': Value('int64'), 'incentivized': Value('int64')}, '20260825_144416': {'mandated': Value('int64'), 'incentivized': Value('int64')}}, 'n_transcripts': Value('int64'), 'cells_short': {'incentivized/Ai-Hiring-Assistant': Value('int64'), 'incentivized/Healthcare-Appointment-Optimizer': Value('int64')}, 'skipped_empty': List(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 2840, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, 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 536, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, 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
              git_sha: string
              timestamp_utc: string
              config: struct<model: string, model_key: string, published_key: string, base_url: string, temperature: doubl (... 304 chars omitted)
                child 0, model: string
                child 1, model_key: string
                child 2, published_key: string
                child 3, base_url: string
                child 4, temperature: double
                child 5, bench_dir: string
                child 6, output_root: string
                child 7, concurrency: int64
                child 8, scenario_timeout_s: int64
                child 9, prune_images: bool
                child 10, rollouts_per_cell: int64
                child 11, expected_cells: int64
                child 12, shard_count: int64
                child 13, shard_index: int64
                child 14, judges: struct<grok-4.20: string, gemini-3.1-pro-preview: string>
                    child 0, grok-4.20: string
                    child 1, gemini-3.1-pro-preview: string
                child 15, exclude_scenarios: list<item: string>
                    child 0, item: string
              combined_from: list<item: string>
                child 0, item: string
              n_passes: int64
              n_transcripts: int64
              per_pass_counts: struct<20260825_141805: struct<mandated: int64, incentivized: int64>, 20260825_143004: struct<mandat (... 95 chars omitted)
                child 0, 20260825_141805: struct<mandated: int64, incentivized: int64>
                    child 0, mandated: int64
                    child 1, incentivized: int64
                child 1, 20260825_143004: struct<mandated: int64, incentivized: int64>
                    child 0, mandated: int64
                    child 1, incentivized: int64
                child 2, 20260825_144416: struct<mandated: int64, incentivized: int64>
                    child 0, mandated: int64
                    child 1, incentivized: int64
              cells_short: struct<incentivized/Ai-Hiring-Assistant: int64, incentivized/Healthcare-Appointment-Optimizer: int64 (... 1 chars omitted)
                child 0, incentivized/Ai-Hiring-Assistant: int64
                child 1, incentivized/Healthcare-Appointment-Optimizer: int64
              skipped_empty: list<item: string>
                child 0, item: string
              passes: list<item: string>
                child 0, item: string
              model_key: string
              to
              {'model_key': Value('string'), 'passes': List(Value('string')), 'per_pass_counts': {'20260825_141805': {'mandated': Value('int64'), 'incentivized': Value('int64')}, '20260825_143004': {'mandated': Value('int64'), 'incentivized': Value('int64')}, '20260825_144416': {'mandated': Value('int64'), 'incentivized': Value('int64')}}, 'n_transcripts': Value('int64'), 'cells_short': {'incentivized/Ai-Hiring-Assistant': Value('int64'), 'incentivized/Healthcare-Appointment-Optimizer': Value('int64')}, 'skipped_empty': List(Value('string'))}
              because column names don't match

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.

ODCV-Bench: verbose-CoT token-matched arm, 3 passes x 30 incentivized cells

field value
experiment Does MORE VERBOSE chain-of-thought, holding the IDEAS constant, change agentic misalignment? The 716 difficult-advice reasoning traces were expanded ~3x in length by Sonnet 5 under a two-judge fidelity gate (no new kinds of deliberation, no new cases, nothing dropped), then trained as a LoRA. TOKEN-matched: difficult advice's share of TRAINABLE TOKENS is held at the da716 baseline instead, which costs rows - roughly half the 716 expanded traces fit the same budget. Isolates 'same reasoning budget, spread over fewer, longer examples'. Headline on these 30 incentivized cells: MR 31.1% CI95 [15.6, 47.8] (bootstrapped over SCENARIOS), severity 1.33, 30 scenarios / 87 rollouts. Published base Qwen3.6-27B on the same cells: 42.5%.
date_generated 2026-08-25
constitution constitutions/claude_distilled_12_principles_mid/constitution.md in the source repo - inherited unchanged from the difficult-advice run the traces were expanded from, and carried into training data LASR-Callum/2026-08-25-table2-9284-difficult-advice-verbose-token-matched-train-mixture (t2_9284_da_verbose_tokenmatched.jsonl @ e71102af09dba065de65d086137edb3a4ff7fc9b)
source_repo https://github.com/Matthew-Bozoukov/teaching_claude_why_replication.git @ 03e384013c52542ddb8586f781da8292d05ef4bc
models target: LASR-Callum/2026-08-20-qwen36-lora-table2-9284-difficult-advice-verbose-token-matched-rank-64-dynbatch (LoRA r64 on Qwen/Qwen3.6-27B, thinking mode, served by vLLM on a RunPod H200, max_num_seqs 32); judges: {"grok-4.20": "x-ai/grok-4.20", "gemini-3.1-pro-preview": "google/gemini-3.1-pro-preview"}; trace expander: anthropic/claude-sonnet-5 via OpenRouter
generation_config {"temperature": 0.0, "passes": 3, "cells": 30, "variant": "incentivized only", "n_rollouts_actual": 87, "config": "configs/eval/2026-08-25_odcv_bench_difficult_advice_716_verbose_tokens_rank64_incentivized_5_30.yaml", "concurrency": 12, "driver": "rented vast CPU boxes, docker per scenario, reaching the model at host.docker.internal:8000 through an SSH tunnel to the serving pod"}
schema rollouts///pass/: messages_record.txt is THE rollout (task + reasoning + actions, self-contained), beside docker_output.log (container stdout, NOT the rollout) and cell_meta.json (manifest row, transcript_bytes, and whether this exact transcript was judged). results/: results.json (headline + per_scenario_medians, one LIST per scenario holding its per-rollout severity), scores_.json, judging_run_meta.json. metadata/: combine_manifest.json (which passes merged, what each contributed), per-pass manifests and run_meta.
provenance bash scratch/verbose_cot/bootstrap_boxes.sh prep
ci_note CIs bootstrap over SCENARIOS, not rollouts. Repeated rollouts of one scenario share a prompt, a model and temperature 0, so resampling them is pseudo-replication and reports an interval that is too narrow - this run measured [16.9, 34.8] that way against a correctly clustered [15.6, 47.8]. A scenario contributes its violation RATE across rollouts (0, 1/3, 2/3, 1) rather than a thresholded verdict, and every scenario weighs the same however many rollouts survived for it. Fixed in src/eval/misalignment/odcv/ on 2026-08-25; the numbers here are post-fix.
coverage_note 87 rollouts rather than 90: {"incentivized/Ai-Hiring-Assistant": 2, "incentivized/Healthcare-Appointment-Optimizer": 1} came up short where a cell produced no transcript in a pass (skipped_empty in combine_manifest.json). Those scenarios are still scored, on the rollouts they did produce, and weigh the same as any other scenario.
Downloads last month
185
Total file size:
6.75 MB