Datasets:

prism-drift
/

reinforcement-learning-results

Tags:

Dataset Viewer

Split (1)

train

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
prompt_idx: int64
anchor_id: string
anchor_task: string
anchor_language_id: null
anchor_dataset_status: string
anchor_example_datapoints_source: string
concreteness_level: string
target_id: string
target_task: string
target_language_id: string
target_dataset_status: string
prompt: string
anchor_training_method: string
anchor_source_cohort_id: string
anchor_example_datapoints_provenance: struct<schema: string, source_kind: string, selection_seed: int64, selected_record_ids: list<item: s (... 3954 chars omitted)
  child 0, schema: string
  child 1, source_kind: string
  child 2, selection_seed: int64
  child 3, selected_record_ids: list<item: string>
      child 0, item: string
  child 4, source_artifacts: list<item: struct<path: string, sha256: string>>
      child 0, item: struct<path: string, sha256: string>
          child 0, path: string
          child 1, sha256: string
  child 5, training_coverage: struct<locked_prompt_rows: int64, model_facing_unique_prompts: int64, not_exposure_matched: bool>
      child 0, locked_prompt_rows: int64
      child 1, model_facing_unique_prompts: int64
      child 2, not_exposure_matched: bool
  child 6, run_metadata: struct<schema: string, status: string, algo: string, base_model: string, dataset: string, dataset_sh (... 3415 chars omitted)
      child 0, schema: string
      child 1, status: string
      child 2, algo: string
      child 3, base_model: string
      child 4, dataset: string
      child 5, dataset_sha256: string
      chil
...
 child 6, qwen35-9b-rl-grpo-s0-step125-prompts-repaired-v1-aws: string
  child 7, qwen35-9b-rl-ppo-s0-step125-prompts-repaired-v1-aws: string
excluded: struct<aggregate_analysis_reports: bool, corrupted_v12_results: bool, legacy_fulltests_v2_results: b (... 110 chars omitted)
  child 0, aggregate_analysis_reports: bool
  child 1, corrupted_v12_results: bool
  child 2, legacy_fulltests_v2_results: bool
  child 3, superseded_method_stimulus_results: bool
  child 4, checkpoints_and_training_data: bool
  child 5, smoke_runs_and_logs: bool
status: string
repo_type: string
training_exposure: struct<locked_prompts_per_size: int64, dpo_four_b_verifier_separable_pairs: int64, dpo_nine_b_verifi (... 60 chars omitted)
  child 0, locked_prompts_per_size: int64
  child 1, dpo_four_b_verifier_separable_pairs: int64
  child 2, dpo_nine_b_verifier_separable_pairs: int64
  child 3, algorithm_exposure_matched: bool
prompt_source: struct<repo_id: string, revision: string, four_b_prompt_hash_prefix: string, nine_b_prompt_hash_pref (... 11 chars omitted)
  child 0, repo_id: string
  child 1, revision: string
  child 2, four_b_prompt_hash_prefix: string
  child 3, nine_b_prompt_hash_prefix: string
release_id: string
created_at: timestamp[s]
inventory_before_release_metadata: struct<result_files: int64, result_bytes: int64, run_state_files: int64, run_state_bytes: int64>
  child 0, result_files: int64
  child 1, result_bytes: int64
  child 2, run_state_files: int64
  child 3, run_state_bytes: int64
to
{'schema': Value('string'), 'repo_id': Value('string'), 'repo_type': Value('string'), 'visibility': Value('string'), 'release_id': Value('string'), 'created_at': Value('timestamp[s]'), 'source_git_commit': Value('string'), 'experiment_namespace': Value('string'), 'status': Value('string'), 'prompt_source': {'repo_id': Value('string'), 'revision': Value('string'), 'four_b_prompt_hash_prefix': Value('string'), 'nine_b_prompt_hash_prefix': Value('string')}, 'scope': {'model_sizes': List(Value('string')), 'algorithms': List(Value('string')), 'checkpoint_step': Value('int64'), 'seed': Value('int64'), 'thinking_modes': List(Value('string')), 'actual_preference_batteries': List(Value('string')), 'anticipation_batteries': List(Value('string')), 'actual_preference_tasks': Value('int64'), 'anticipation_tasks': Value('int64'), 'anticipation_forecasts': Value('int64'), 'anticipation_raw_responses': Value('int64'), 'anticipation_valid_responses': Value('int64'), 'four_b_anticipation_valid_rate': Value('float64'), 'nine_b_anticipation_valid_rate': Value('float64')}, 'model_artifact_identities': {'qwen35-4b-m0-v4-prompts-repaired-v1-aws': Value('string'), 'qwen35-4b-rl-dpo-s0-step125-prompts-repaired-v1-aws': Value('string'), 'qwen35-4b-rl-grpo-s0-step125-prompts-repaired-v1-aws': Value('string'), 'qwen35-4b-rl-ppo-s0-step125-prompts-repaired-v1-aws': Value('string'), 'qwen35-9b-m0-v4-prompts-repaired-v1-aws': Value('string'), 'qwen35-9b-rl-dpo-s0-step125-prompts-repaired-v1-aws': Value('string'), 'qwen35-9b-rl-grpo-s0-step125-prompts-repaired-v1-aws': Value('string'), 'qwen35-9b-rl-ppo-s0-step125-prompts-repaired-v1-aws': Value('string')}, 'inventory_before_release_metadata': {'result_files': Value('int64'), 'result_bytes': Value('int64'), 'run_state_files': Value('int64'), 'run_state_bytes': Value('int64')}, 'training_exposure': {'locked_prompts_per_size': Value('int64'), 'dpo_four_b_verifier_separable_pairs': Value('int64'), 'dpo_nine_b_verifier_separable_pairs': Value('int64'), 'algorithm_exposure_matched': Value('bool')}, 'excluded': {'aggregate_analysis_reports': Value('bool'), 'corrupted_v12_results': Value('bool'), 'legacy_fulltests_v2_results': Value('bool'), 'superseded_method_stimulus_results': Value('bool'), 'checkpoints_and_training_data': Value('bool'), 'smoke_runs_and_logs': Value('bool')}}
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
              prompt_idx: int64
              anchor_id: string
              anchor_task: string
              anchor_language_id: null
              anchor_dataset_status: string
              anchor_example_datapoints_source: string
              concreteness_level: string
              target_id: string
              target_task: string
              target_language_id: string
              target_dataset_status: string
              prompt: string
              anchor_training_method: string
              anchor_source_cohort_id: string
              anchor_example_datapoints_provenance: struct<schema: string, source_kind: string, selection_seed: int64, selected_record_ids: list<item: s (... 3954 chars omitted)
                child 0, schema: string
                child 1, source_kind: string
                child 2, selection_seed: int64
                child 3, selected_record_ids: list<item: string>
                    child 0, item: string
                child 4, source_artifacts: list<item: struct<path: string, sha256: string>>
                    child 0, item: struct<path: string, sha256: string>
                        child 0, path: string
                        child 1, sha256: string
                child 5, training_coverage: struct<locked_prompt_rows: int64, model_facing_unique_prompts: int64, not_exposure_matched: bool>
                    child 0, locked_prompt_rows: int64
                    child 1, model_facing_unique_prompts: int64
                    child 2, not_exposure_matched: bool
                child 6, run_metadata: struct<schema: string, status: string, algo: string, base_model: string, dataset: string, dataset_sh (... 3415 chars omitted)
                    child 0, schema: string
                    child 1, status: string
                    child 2, algo: string
                    child 3, base_model: string
                    child 4, dataset: string
                    child 5, dataset_sha256: string
                    chil
              ...
               child 6, qwen35-9b-rl-grpo-s0-step125-prompts-repaired-v1-aws: string
                child 7, qwen35-9b-rl-ppo-s0-step125-prompts-repaired-v1-aws: string
              excluded: struct<aggregate_analysis_reports: bool, corrupted_v12_results: bool, legacy_fulltests_v2_results: b (... 110 chars omitted)
                child 0, aggregate_analysis_reports: bool
                child 1, corrupted_v12_results: bool
                child 2, legacy_fulltests_v2_results: bool
                child 3, superseded_method_stimulus_results: bool
                child 4, checkpoints_and_training_data: bool
                child 5, smoke_runs_and_logs: bool
              status: string
              repo_type: string
              training_exposure: struct<locked_prompts_per_size: int64, dpo_four_b_verifier_separable_pairs: int64, dpo_nine_b_verifi (... 60 chars omitted)
                child 0, locked_prompts_per_size: int64
                child 1, dpo_four_b_verifier_separable_pairs: int64
                child 2, dpo_nine_b_verifier_separable_pairs: int64
                child 3, algorithm_exposure_matched: bool
              prompt_source: struct<repo_id: string, revision: string, four_b_prompt_hash_prefix: string, nine_b_prompt_hash_pref (... 11 chars omitted)
                child 0, repo_id: string
                child 1, revision: string
                child 2, four_b_prompt_hash_prefix: string
                child 3, nine_b_prompt_hash_prefix: string
              release_id: string
              created_at: timestamp[s]
              inventory_before_release_metadata: struct<result_files: int64, result_bytes: int64, run_state_files: int64, run_state_bytes: int64>
                child 0, result_files: int64
                child 1, result_bytes: int64
                child 2, run_state_files: int64
                child 3, run_state_bytes: int64
              to
              {'schema': Value('string'), 'repo_id': Value('string'), 'repo_type': Value('string'), 'visibility': Value('string'), 'release_id': Value('string'), 'created_at': Value('timestamp[s]'), 'source_git_commit': Value('string'), 'experiment_namespace': Value('string'), 'status': Value('string'), 'prompt_source': {'repo_id': Value('string'), 'revision': Value('string'), 'four_b_prompt_hash_prefix': Value('string'), 'nine_b_prompt_hash_prefix': Value('string')}, 'scope': {'model_sizes': List(Value('string')), 'algorithms': List(Value('string')), 'checkpoint_step': Value('int64'), 'seed': Value('int64'), 'thinking_modes': List(Value('string')), 'actual_preference_batteries': List(Value('string')), 'anticipation_batteries': List(Value('string')), 'actual_preference_tasks': Value('int64'), 'anticipation_tasks': Value('int64'), 'anticipation_forecasts': Value('int64'), 'anticipation_raw_responses': Value('int64'), 'anticipation_valid_responses': Value('int64'), 'four_b_anticipation_valid_rate': Value('float64'), 'nine_b_anticipation_valid_rate': Value('float64')}, 'model_artifact_identities': {'qwen35-4b-m0-v4-prompts-repaired-v1-aws': Value('string'), 'qwen35-4b-rl-dpo-s0-step125-prompts-repaired-v1-aws': Value('string'), 'qwen35-4b-rl-grpo-s0-step125-prompts-repaired-v1-aws': Value('string'), 'qwen35-4b-rl-ppo-s0-step125-prompts-repaired-v1-aws': Value('string'), 'qwen35-9b-m0-v4-prompts-repaired-v1-aws': Value('string'), 'qwen35-9b-rl-dpo-s0-step125-prompts-repaired-v1-aws': Value('string'), 'qwen35-9b-rl-grpo-s0-step125-prompts-repaired-v1-aws': Value('string'), 'qwen35-9b-rl-ppo-s0-step125-prompts-repaired-v1-aws': Value('string')}, 'inventory_before_release_metadata': {'result_files': Value('int64'), 'result_bytes': Value('int64'), 'run_state_files': Value('int64'), 'run_state_bytes': Value('int64')}, 'training_exposure': {'locked_prompts_per_size': Value('int64'), 'dpo_four_b_verifier_separable_pairs': Value('int64'), 'dpo_nine_b_verifier_separable_pairs': Value('int64'), 'algorithm_exposure_matched': Value('bool')}, 'excluded': {'aggregate_analysis_reports': Value('bool'), 'corrupted_v12_results': Value('bool'), 'legacy_fulltests_v2_results': Value('bool'), 'superseded_method_stimulus_results': Value('bool'), 'checkpoints_and_training_data': Value('bool'), 'smoke_runs_and_logs': Value('bool')}}
              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.

Reinforcement Learning Results

This repository contains the latest result artifacts for the prompts-repaired-v1 reinforcement-learning experiments in ai-pref-drift.

Included scope

Model sizes: Qwen3.5 4B and 9B.
Model states: M0-v4 and DPO, GRPO, and PPO Step 125 (seed 0).
Actual pairwise-preference batteries: coding-task preference and value preference, each in thinking and non-thinking modes.
Anticipation batteries: coding-task anticipation and value anticipation, each in thinking and non-thinking modes.
Parsed results, raw responses, fitted utilities, text summaries, elicitation manifests, completion markers, run plans, and immutable model artifact identities.

The result files preserve their original local experiment-directory names so that model, algorithm, seed, checkpoint, and prompt-cohort provenance remain explicit.

Intentionally excluded

Aggregate analysis reports. Analysis is maintained separately and was not uploaded in this release.
Results derived from the corrupted v12 or legacy fulltests-v2 cohorts.
Superseded anticipation and described-preference runs whose provenance referenced those legacy artifacts.
Model checkpoints, reward models, training datasets, smoke runs, and logs.

Directory layout

results/<model-key>/anticipation/   Latest repaired-v1 anticipation artifacts
results/<model-key>/pairs/          Actual coding-task preference artifacts
results/<model-key>/values/         Actual value-preference artifacts
run_state/actual_preference/        Plans, completion markers, identities
run_state/anticipation/             Plans, completion markers, identities
SHA256SUMS                           Checksums for every uploaded data artifact
UPLOAD_MANIFEST.json                Release-level inventory and provenance

Important limitations

Status is provisional_verifier_v1: verifier-v2 was deliberately not used for this release.
There is one training seed, so these results do not establish a population- level causal effect of algorithm choice.
DPO used only verifier-separable pairs (724 for 4B and 739 for 9B), while GRPO and PPO used the locked 1,000-prompt cohort. Algorithm choice and training-data exposure are therefore partially confounded.
Anticipation was collected after training completed. M0 outputs are baseline-state counterfactual forecasts, not temporally preregistered prospective predictions.

Every uploaded result belongs to the prompts-repaired-v1 namespace. See UPLOAD_MANIFEST.json and SHA256SUMS for the exact release contents.

Downloads last month: 15

Total file size:

2.26 GB