The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
model: string
benchmark: string
test_file: string
num_prompts: int64
num_shards: int64
shard_index: int64
n: int64
ks: list<item: int64>
child 0, item: int64
temperature: double
top_p: double
top_k: int64
max_new_tokens: int64
pass@1: double
pass@1_by_difficulty: struct<0/8: struct<pak: double, n: int64>, 1/8: struct<pak: double, n: int64>, 2/8: struct<pak: doub (... 194 chars omitted)
child 0, 0/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 1, 1/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 2, 2/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 3, 3/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 4, 4/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 5, 5/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 6, 6/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 7, 7/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
pass@2: double
pass@2_by_difficulty: struct<0/8: struct<pak: double, n: int64>, 1/8: struct<pak: double, n: int64>, 2/8: struct<pak: doub (... 194 chars omitted)
child 0, 0/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 1, 1/8: struct<pak: double, n: int64>
child 0, pak: double
...
child 1, n: int64
child 3, 3/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 4, 4/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 5, 5/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 6, 6/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 7, 7/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
pass@128: double
pass@128_by_difficulty: struct<0/8: struct<pak: double, n: int64>, 1/8: struct<pak: double, n: int64>, 2/8: struct<pak: doub (... 194 chars omitted)
child 0, 0/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 1, 1/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 2, 2/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 3, 3/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 4, 4/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 5, 5/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 6, 6/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 7, 7/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
idx: int64
difficulty: string
num_correct: int64
to
{'idx': Value('int64'), 'difficulty': Value('string'), 'num_correct': Value('int64'), 'n': Value('int64')}
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
model: string
benchmark: string
test_file: string
num_prompts: int64
num_shards: int64
shard_index: int64
n: int64
ks: list<item: int64>
child 0, item: int64
temperature: double
top_p: double
top_k: int64
max_new_tokens: int64
pass@1: double
pass@1_by_difficulty: struct<0/8: struct<pak: double, n: int64>, 1/8: struct<pak: double, n: int64>, 2/8: struct<pak: doub (... 194 chars omitted)
child 0, 0/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 1, 1/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 2, 2/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 3, 3/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 4, 4/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 5, 5/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 6, 6/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 7, 7/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
pass@2: double
pass@2_by_difficulty: struct<0/8: struct<pak: double, n: int64>, 1/8: struct<pak: double, n: int64>, 2/8: struct<pak: doub (... 194 chars omitted)
child 0, 0/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 1, 1/8: struct<pak: double, n: int64>
child 0, pak: double
...
child 1, n: int64
child 3, 3/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 4, 4/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 5, 5/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 6, 6/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 7, 7/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
pass@128: double
pass@128_by_difficulty: struct<0/8: struct<pak: double, n: int64>, 1/8: struct<pak: double, n: int64>, 2/8: struct<pak: doub (... 194 chars omitted)
child 0, 0/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 1, 1/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 2, 2/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 3, 3/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 4, 4/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 5, 5/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 6, 6/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
child 7, 7/8: struct<pak: double, n: int64>
child 0, pak: double
child 1, n: int64
idx: int64
difficulty: string
num_correct: int64
to
{'idx': Value('int64'), 'difficulty': Value('string'), 'num_correct': Value('int64'), 'n': Value('int64')}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Polaris 8-GPU experiment results
All four training runs are complete at step 4000, with 142 unique evaluation points (1136 shards). Both Qwen3-8B-Base runs were extended from their full step-2000 checkpoints to step 4000. Both Llama-3.2-3B-Instruct runs finished at step 4000. Each model has a no-replay arm and a hard-cooldown replay arm with lambda 0.1.
Evaluation
Each point covers all 800 prompts with 160 samples per prompt. The shared base
model is step 0. Original runs were evaluated every 100 steps; the Qwen extensions
are evaluated at steps 2200, 2400, ..., 4000, every 200 steps. All 20 extension points are complete. Sampling uses
temperature 0.6, top-p 0.95, top-k -1, and at most 3072 generated tokens.
tables/all_pass_at_k.csv contains all points, including both base models.
Scores are percentages, computed with the unbiased pass@k estimator.
Use sub600 (shards 0,1,3,4,5,7) when comparing with existing Polaris curves.
full800 uses every prompt. The four other CSVs preserve the original merged tables.
Raw summaries and per-prompt correctness counts are under evals/.
Archives and provenance
archives/ contains compressed training rollout data, replayed rollout data,
logs, W&B run files, and run metadata from the original experiment (Qwen through
2000 and Llama through 4000). Extension evaluation results are included; extension
rollout/log archives are not part of this update. Archive paths match the original Modal
volume layout; extract into an empty directory with tar -xzf <archive>.
JSONL manifests in manifests/ record every source file and its SHA-256.
SHA256SUMS covers the uploaded artifacts. provenance/source/ contains the
launcher, instructions, shared scripts, and patched trainer used for this experiment.
Model weights, optimizer state, checkpoint replay buffers, and the model cache
are not part of this results dataset. They remain on Modal volume
rl-forgetting-polaris-8gpu. Raw measurements there were not modified.
Source: https://github.com/jy-evangeline/rl-forgetting-experiments/tree/polaris-8gpu Upstream commit: 041a6be. Local Modal integration preserves full checkpoints and adds checkpoint-based continuation. Evaluation fixes set the repo import path and validate safetensors payloads instead of trusting an inaccurate total-size field. The original model and dataset licenses continue to apply to derived artifacts.
- Downloads last month
- 105