Hugging Face's logo

Datasets:
LASR-Callum
/
2026-08-06-model-eval-model-self

Tags:
training-data
kind:synth
pipeline:model-eval-model
constitution:claude_distilled_09_principles_mid_20260804
backfilled-2026-08-25
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
scenario_id: string
trait_id: string
trait_name: string
trait_text: string
domain: string
situation: string
shortcut: string
draft_system: string
draft_user: string
system: string
user: string
refine_changes: string
draft_reasoning: string
draft_response: string
reasoning: string
response: string
rewrite_changes: string
ablated: list<item: null>
  child 0, item: null
source: struct<source_run: string, source_git_sha: string, source_constitution_sha256: string>
  child 0, source_run: string
  child 1, source_git_sha: string
  child 2, source_constitution_sha256: string
run_dir: string
pipeline: string
config: struct<pipeline: string, seed: int64, constitution: string, source: struct<hf_repo: string>, output_ (... 1826 chars omitted)
  child 0, pipeline: string
  child 1, seed: int64
  child 2, constitution: string
  child 3, source: struct<hf_repo: string>
      child 0, hf_repo: string
  child 4, output_dir: string
  child 5, hf_repo: string
  child 6, hf_repo_smoke: null
  child 7, hf_private: bool
  child 8, workers: int64
  child 9, budget_usd: double
  child 10, smoke: struct<cells: struct<m2_self_good: int64, m1_self_flawed: int64>, hf_repo: null>
      child 0, cells: struct<m2_self_good: int64, m1_self_flawed: int64>
          child 0, m2_self_good: int64
          child 1, m1_self_flawed: int64
      child 1, hf_repo: null
  child 11, cells: struct<m2_self_good: int64, m1_self_flawed: int64>
      child 0, m2_self_good: int64
      child 1, m1_self_flawed: int64
  chi
...
: string
      child 6, reflect_system: string
      child 7, reflect_format: string
effective: struct<>
constitution_sha256: string
wall_clock_s: double
smoke: bool
stage_seconds: struct<generated: double, sft: double>
  child 0, generated: double
  child 1, sft: double
hf_repo: string
run_id: string
git_sha: string
counts: struct<source: int64, plan: int64, perturbed: int64, generated: int64, sft: int64>
  child 0, source: int64
  child 1, plan: int64
  child 2, perturbed: int64
  child 3, generated: int64
  child 4, sft: int64
usage: struct<by_model: struct<anthropic/claude-sonnet-5: struct<calls: int64, prompt_tokens: int64, comple (... 163 chars omitted)
  child 0, by_model: struct<anthropic/claude-sonnet-5: struct<calls: int64, prompt_tokens: int64, completion_tokens: int6 (... 16 chars omitted)
      child 0, anthropic/claude-sonnet-5: struct<calls: int64, prompt_tokens: int64, completion_tokens: int64, usd: double>
          child 0, calls: int64
          child 1, prompt_tokens: int64
          child 2, completion_tokens: int64
          child 3, usd: double
  child 1, by_stage: struct<reflect: struct<calls: int64, prompt_tokens: int64, completion_tokens: int64, usd: double>>
      child 0, reflect: struct<calls: int64, prompt_tokens: int64, completion_tokens: int64, usd: double>
          child 0, calls: int64
          child 1, prompt_tokens: int64
          child 2, completion_tokens: int64
          child 3, usd: double
  child 2, total_usd: double
workers: int64
to
{'run_id': Value('string'), 'pipeline': Value('string'), 'git_sha': Value('string'), 'smoke': Value('bool'), 'constitution_sha256': Value('string'), 'config': {'pipeline': Value('string'), 'seed': Value('int64'), 'constitution': Value('string'), 'source': {'hf_repo': Value('string')}, 'output_dir': Value('string'), 'hf_repo': Value('string'), 'hf_repo_smoke': Value('null'), 'hf_private': Value('bool'), 'workers': Value('int64'), 'budget_usd': Value('float64'), 'smoke': {'cells': {'m2_self_good': Value('int64'), 'm1_self_flawed': Value('int64')}, 'hf_repo': Value('null')}, 'cells': {'m2_self_good': Value('int64'), 'm1_self_flawed': Value('int64')}, 'explicitness': {'name_clause': Value('float64'), 'paraphrase': Value('float64'), 'embody': Value('float64')}, 'flaws': {'types': {'omission': Value('float64'), 'commission': Value('float64'), 'miscalibration': Value('float64'), 'over_application': Value('float64')}, 'severities': {'clear': Value('float64'), 'moderate': Value('float64')}}, 'defaults': {'temperature': Value('float64'), 'max_tokens': Value('int64')}, 'models': {'reflect': {'model': Value('string'), 'temperature': Value('float64'), 'max_tokens': Value('int64'), 'reasoning': {'enabled': Value('bool')}, 'assumed_tokens': {'in': Value('int64'), 'out': Value('int64')}}, 'perturb': {'model': Value('string'), 'temperature': Value('float64'), 'max_tokens': Value('int64'), 'reasoning': {'enabled': Value('bool')}, 'assumed_tokens': {'in': Value('int64'), 'out': Value('int64')}}
...
int': Value('string'), 'prompts': {'system': Value('string'), 'user': Value('string')}}), 'prompts': {'explicitness_styles': {'name_clause': Value('string'), 'paraphrase': Value('string'), 'embody': Value('string')}, 'flaw_types': {'omission': Value('string'), 'commission': Value('string'), 'miscalibration': Value('string'), 'over_application': Value('string')}, 'flaw_severities': {'clear': Value('string'), 'moderate': Value('string')}, 'transcript_wrappers': List(Value('string')), 'reflect_variants': List(Value('string')), 'known_flaw_note': Value('string'), 'reflect_system': Value('string'), 'reflect_format': Value('string')}}, 'effective': {}, 'ablated': List(Value('null')), 'counts': {'source': Value('int64'), 'plan': Value('int64'), 'perturbed': Value('int64'), 'generated': Value('int64'), 'sft': Value('int64')}, 'usage': {'by_model': {'anthropic/claude-sonnet-5': {'calls': Value('int64'), 'prompt_tokens': Value('int64'), 'completion_tokens': Value('int64'), 'usd': Value('float64')}}, 'by_stage': {'reflect': {'calls': Value('int64'), 'prompt_tokens': Value('int64'), 'completion_tokens': Value('int64'), 'usd': Value('float64')}}, 'total_usd': Value('float64')}, 'wall_clock_s': Value('float64'), 'stage_seconds': {'generated': Value('float64'), 'sft': Value('float64')}, 'workers': Value('int64'), 'hf_repo': Value('string'), 'run_dir': Value('string'), 'source': {'source_run': Value('string'), 'source_git_sha': Value('string'), 'source_constitution_sha256': 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 489, 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 2818, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2355, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2380, 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 2369, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2297, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              scenario_id: string
              trait_id: string
              trait_name: string
              trait_text: string
              domain: string
              situation: string
              shortcut: string
              draft_system: string
              draft_user: string
              system: string
              user: string
              refine_changes: string
              draft_reasoning: string
              draft_response: string
              reasoning: string
              response: string
              rewrite_changes: string
              ablated: list<item: null>
                child 0, item: null
              source: struct<source_run: string, source_git_sha: string, source_constitution_sha256: string>
                child 0, source_run: string
                child 1, source_git_sha: string
                child 2, source_constitution_sha256: string
              run_dir: string
              pipeline: string
              config: struct<pipeline: string, seed: int64, constitution: string, source: struct<hf_repo: string>, output_ (... 1826 chars omitted)
                child 0, pipeline: string
                child 1, seed: int64
                child 2, constitution: string
                child 3, source: struct<hf_repo: string>
                    child 0, hf_repo: string
                child 4, output_dir: string
                child 5, hf_repo: string
                child 6, hf_repo_smoke: null
                child 7, hf_private: bool
                child 8, workers: int64
                child 9, budget_usd: double
                child 10, smoke: struct<cells: struct<m2_self_good: int64, m1_self_flawed: int64>, hf_repo: null>
                    child 0, cells: struct<m2_self_good: int64, m1_self_flawed: int64>
                        child 0, m2_self_good: int64
                        child 1, m1_self_flawed: int64
                    child 1, hf_repo: null
                child 11, cells: struct<m2_self_good: int64, m1_self_flawed: int64>
                    child 0, m2_self_good: int64
                    child 1, m1_self_flawed: int64
                chi
              ...
              : string
                    child 6, reflect_system: string
                    child 7, reflect_format: string
              effective: struct<>
              constitution_sha256: string
              wall_clock_s: double
              smoke: bool
              stage_seconds: struct<generated: double, sft: double>
                child 0, generated: double
                child 1, sft: double
              hf_repo: string
              run_id: string
              git_sha: string
              counts: struct<source: int64, plan: int64, perturbed: int64, generated: int64, sft: int64>
                child 0, source: int64
                child 1, plan: int64
                child 2, perturbed: int64
                child 3, generated: int64
                child 4, sft: int64
              usage: struct<by_model: struct<anthropic/claude-sonnet-5: struct<calls: int64, prompt_tokens: int64, comple (... 163 chars omitted)
                child 0, by_model: struct<anthropic/claude-sonnet-5: struct<calls: int64, prompt_tokens: int64, completion_tokens: int6 (... 16 chars omitted)
                    child 0, anthropic/claude-sonnet-5: struct<calls: int64, prompt_tokens: int64, completion_tokens: int64, usd: double>
                        child 0, calls: int64
                        child 1, prompt_tokens: int64
                        child 2, completion_tokens: int64
                        child 3, usd: double
                child 1, by_stage: struct<reflect: struct<calls: int64, prompt_tokens: int64, completion_tokens: int64, usd: double>>
                    child 0, reflect: struct<calls: int64, prompt_tokens: int64, completion_tokens: int64, usd: double>
                        child 0, calls: int64
                        child 1, prompt_tokens: int64
                        child 2, completion_tokens: int64
                        child 3, usd: double
                child 2, total_usd: double
              workers: int64
              to
              {'run_id': Value('string'), 'pipeline': Value('string'), 'git_sha': Value('string'), 'smoke': Value('bool'), 'constitution_sha256': Value('string'), 'config': {'pipeline': Value('string'), 'seed': Value('int64'), 'constitution': Value('string'), 'source': {'hf_repo': Value('string')}, 'output_dir': Value('string'), 'hf_repo': Value('string'), 'hf_repo_smoke': Value('null'), 'hf_private': Value('bool'), 'workers': Value('int64'), 'budget_usd': Value('float64'), 'smoke': {'cells': {'m2_self_good': Value('int64'), 'm1_self_flawed': Value('int64')}, 'hf_repo': Value('null')}, 'cells': {'m2_self_good': Value('int64'), 'm1_self_flawed': Value('int64')}, 'explicitness': {'name_clause': Value('float64'), 'paraphrase': Value('float64'), 'embody': Value('float64')}, 'flaws': {'types': {'omission': Value('float64'), 'commission': Value('float64'), 'miscalibration': Value('float64'), 'over_application': Value('float64')}, 'severities': {'clear': Value('float64'), 'moderate': Value('float64')}}, 'defaults': {'temperature': Value('float64'), 'max_tokens': Value('int64')}, 'models': {'reflect': {'model': Value('string'), 'temperature': Value('float64'), 'max_tokens': Value('int64'), 'reasoning': {'enabled': Value('bool')}, 'assumed_tokens': {'in': Value('int64'), 'out': Value('int64')}}, 'perturb': {'model': Value('string'), 'temperature': Value('float64'), 'max_tokens': Value('int64'), 'reasoning': {'enabled': Value('bool')}, 'assumed_tokens': {'in': Value('int64'), 'out': Value('int64')}}
              ...
              int': Value('string'), 'prompts': {'system': Value('string'), 'user': Value('string')}}), 'prompts': {'explicitness_styles': {'name_clause': Value('string'), 'paraphrase': Value('string'), 'embody': Value('string')}, 'flaw_types': {'omission': Value('string'), 'commission': Value('string'), 'miscalibration': Value('string'), 'over_application': Value('string')}, 'flaw_severities': {'clear': Value('string'), 'moderate': Value('string')}, 'transcript_wrappers': List(Value('string')), 'reflect_variants': List(Value('string')), 'known_flaw_note': Value('string'), 'reflect_system': Value('string'), 'reflect_format': Value('string')}}, 'effective': {}, 'ablated': List(Value('null')), 'counts': {'source': Value('int64'), 'plan': Value('int64'), 'perturbed': Value('int64'), 'generated': Value('int64'), 'sft': Value('int64')}, 'usage': {'by_model': {'anthropic/claude-sonnet-5': {'calls': Value('int64'), 'prompt_tokens': Value('int64'), 'completion_tokens': Value('int64'), 'usd': Value('float64')}}, 'by_stage': {'reflect': {'calls': Value('int64'), 'prompt_tokens': Value('int64'), 'completion_tokens': Value('int64'), 'usd': Value('float64')}}, 'total_usd': Value('float64')}, 'wall_clock_s': Value('float64'), 'stage_seconds': {'generated': Value('float64'), 'sft': Value('float64')}, 'workers': Value('int64'), 'hf_repo': Value('string'), 'run_dir': Value('string'), 'source': {'source_run': Value('string'), 'source_git_sha': Value('string'), 'source_constitution_sha256': 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.

2026-08-06-model-eval-model-self

Self-arm corpus of the model-eval-model experiment: 2,087 multi-turn self-evaluation documents (m1_self_flawed / m2_self_good) for SFT of the first self-evaluation model organism.

field value
experiment model-eval-model self-arm: >2-turn self-evaluation documents (user → assistant reply → "look back at it" → supervised reflection turn) teaching "revise when the earlier reply was flawed, hold when it was sound"
date_generated 2026-08-06
constitution claude_distilled_09_principles_mid_20260804 — the recovered 9-principle interim state the source corpus was generated against (sha256 fe2ed960…); see constitutions/claude_distilled_09_principles_mid_20260804/ in the source repo
source_repo https://github.com/LASR-Labs/teaching_claude_why_replication @ 6e5be8b + uncommitted self-arm changes of 2026-08-06 (config configs/data/synthdoc/model_eval_model_self.yaml, known_flaw support in src/data/synthdoc/cells.py; see LOG 2026-08-06)
models generation + perturbation: anthropic/claude-sonnet-5 via OpenRouter (hidden reasoning disabled); check judges: anthropic/claude-sonnet-5
generation_config reflect: temp 0.8, max_tokens 12288; perturb: temp 0.7, max_tokens 8192; seed 0; full settings in the config file above
schema stage_5_sft.jsonl: {messages, metadata} — messages = [system, user, assistant (evaluated reply, no think), user (reflect prompt), assistant (reflection, reasoning_content = think trace)]; metadata.supervise: "final" (only the last turn trains); metadata carries cell, verdict (held/revised), flaw_type/severity (m1 only; never in the text), explicitness, trait, scenario_id
provenance uv run scripts/data/synthdoc/build_dataset.py --config configs/data/synthdoc/model_eval_model_self.yaml over source corpus LASR-Callum/2026-08-04-synthdoc-package-difficult-advice-stage-cache (gold responses verbatim; m1 twins minimally perturbed, one labeled flaw each)

Design notes

  • m1 generation is deliberately unblinded: the perturbation's change_summary is fed to the generator as scaffolding (known_flaw_note) so the reflection reliably rediscovers and fixes the planted flaw. The training text stays blind — the check suite gates on the summary never appearing in any training message (generator_blind: false is reported honestly). m2 is fully blind.
  • Composition: m1 1,037 docs (990 revised / 47 held), m2 1,050 (975 held / 75 revised); flaws omission/commission/miscalibration/over_application × clear/moderate (grey excluded — forcing revision of defensible replies would train capitulation).
  • Checks (checks_report.json in this repo): surface-shortcut AUC 0.565 (max 0.65, shuffled baseline 0.50), flaw-identification 89% clear / 95% moderate, gold validation 1% below 3/5, post-hoc 13%. template_8gram_share_max was raised 0.20→0.30 for this config with documented rationale: the sole offender is a first-sentence opener stem ("let me actually…") inherited from the generator family — the source corpus's traces share the same opener in ~65% of records — while documents are otherwise diverse (pairwise 4-gram jaccard 0.002).
  • Stage snapshots (stage_1–stage_5) mirror the run dir output/model_eval_model_self/20260806_105121; stage_5_sft.jsonl is the training artifact.
Downloads last month
274
Total file size:
155 MB