Fantastic Pretraining Optimizers and Where to Find Them
Paper • 2509.02046 • Published • 15
Dataset Viewer
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
id: string
language: string
source: string
reference: string
input_ids: list<item: int64>
child 0, item: int64
generated_ids: list<item: int64>
child 0, item: int64
generation: string
dtype: string
max_new_tokens: int64
widths: struct<1: struct<prefill_positions: int64, cached_decode_positions: int64, forward_calls: int64, tot (... 371 chars omitted)
child 0, 1: struct<prefill_positions: int64, cached_decode_positions: int64, forward_calls: int64, total_forward (... 50 chars omitted)
child 0, prefill_positions: int64
child 1, cached_decode_positions: int64
child 2, forward_calls: int64
child 3, total_forward_positions: int64
child 4, saved_outputs_reproduced: bool
child 1, 2: struct<prefill_positions: int64, cached_decode_positions: int64, forward_calls: int64, total_forward (... 50 chars omitted)
child 0, prefill_positions: int64
child 1, cached_decode_positions: int64
child 2, forward_calls: int64
child 3, total_forward_positions: int64
child 4, saved_outputs_reproduced: bool
child 2, 4: struct<prefill_positions: int64, cached_decode_positions: int64, forward_calls: int64, total_forward (... 50 chars omitted)
child 0, prefill_positions: int64
child 1, cached_decode_positions: int64
child 2, forward_calls: int64
child 3, total_forward_positions: int64
child 4, saved_outputs_reproduced: bool
batch_size: int64
definition: string
samples: int64
to
{'samples': Value('int64'), 'batch_size': Value('int64'), 'max_new_tokens': Value('int64'), 'dtype': Value('string'), 'definition': Value('string'), 'widths': {'1': {'prefill_positions': Value('int64'), 'cached_decode_positions': Value('int64'), 'forward_calls': Value('int64'), 'total_forward_positions': Value('int64'), 'saved_outputs_reproduced': Value('bool')}, '2': {'prefill_positions': Value('int64'), 'cached_decode_positions': Value('int64'), 'forward_calls': Value('int64'), 'total_forward_positions': Value('int64'), 'saved_outputs_reproduced': Value('bool')}, '4': {'prefill_positions': Value('int64'), 'cached_decode_positions': Value('int64'), 'forward_calls': Value('int64'), 'total_forward_positions': Value('int64'), 'saved_outputs_reproduced': 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 2951, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, 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 547, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, 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
id: string
language: string
source: string
reference: string
input_ids: list<item: int64>
child 0, item: int64
generated_ids: list<item: int64>
child 0, item: int64
generation: string
dtype: string
max_new_tokens: int64
widths: struct<1: struct<prefill_positions: int64, cached_decode_positions: int64, forward_calls: int64, tot (... 371 chars omitted)
child 0, 1: struct<prefill_positions: int64, cached_decode_positions: int64, forward_calls: int64, total_forward (... 50 chars omitted)
child 0, prefill_positions: int64
child 1, cached_decode_positions: int64
child 2, forward_calls: int64
child 3, total_forward_positions: int64
child 4, saved_outputs_reproduced: bool
child 1, 2: struct<prefill_positions: int64, cached_decode_positions: int64, forward_calls: int64, total_forward (... 50 chars omitted)
child 0, prefill_positions: int64
child 1, cached_decode_positions: int64
child 2, forward_calls: int64
child 3, total_forward_positions: int64
child 4, saved_outputs_reproduced: bool
child 2, 4: struct<prefill_positions: int64, cached_decode_positions: int64, forward_calls: int64, total_forward (... 50 chars omitted)
child 0, prefill_positions: int64
child 1, cached_decode_positions: int64
child 2, forward_calls: int64
child 3, total_forward_positions: int64
child 4, saved_outputs_reproduced: bool
batch_size: int64
definition: string
samples: int64
to
{'samples': Value('int64'), 'batch_size': Value('int64'), 'max_new_tokens': Value('int64'), 'dtype': Value('string'), 'definition': Value('string'), 'widths': {'1': {'prefill_positions': Value('int64'), 'cached_decode_positions': Value('int64'), 'forward_calls': Value('int64'), 'total_forward_positions': Value('int64'), 'saved_outputs_reproduced': Value('bool')}, '2': {'prefill_positions': Value('int64'), 'cached_decode_positions': Value('int64'), 'forward_calls': Value('int64'), 'total_forward_positions': Value('int64'), 'saved_outputs_reproduced': Value('bool')}, '4': {'prefill_positions': Value('int64'), 'cached_decode_positions': Value('int64'), 'forward_calls': Value('int64'), 'total_forward_positions': Value('int64'), 'saved_outputs_reproduced': Value('bool')}}}
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.
Advanced NLP Assignment 2
Pratyush Jena · Monsoon 2026
Five MoE translation variants, four custom pretraining optimizers, a Lion sweep, and twelve forward-only Pythia-160M decoding configurations. The final report contains eight content pages plus one references page.
Read the results
- Report PDF: methods, comparisons, per-layer routing, uncertainty, examples and artifact links.
- Report source and complete measured tables.
- Publication record, checkpoint audit, and W&B audit.
- GPU verification: no cross-example answer leakage; FP32 greedy batching matches on 32 checked prompts; full-reference PPL 25.036.
Reproduce
conda activate moe
uv sync --frozen
# Use the existing moe interpreter with dependencies from the uv lock:
uv export --frozen --no-hashes --format requirements-txt -o /tmp/a2-requirements.txt
uv pip install --python "$CONDA_PREFIX/bin/python" -r /tmp/a2-requirements.txt
scripts/run.sh prepare
scripts/run.sh part1
scripts/run.sh part2
scripts/run.sh part3
scripts/run.sh benchmark
python scripts/measure_beam_work.py # CUDA required
python scripts/collect_routing_controls.py # CUDA required
scripts/run.sh report
python scripts/audit_checkpoints.py
python scripts/audit_wandb.py
python -m pytest -q
run.sh uses conda moe and the existing Slurm allocation 480. On another machine, run python main.py ... with your own CUDA allocation. The reruns' recorded executable was .venv/bin/python; package versions and source hashes are recorded in each provenance.json. uv.lock is the committed dependency source. All runs default to online W&B; --smoke keeps shortened runs separate. Interrupted training resumes model, optimizer, RNG, data cursor and run identity from latest.pt. Rerunning a completed training command skips it; archive its directory deliberately before starting another experiment. Full settings: configs/full.yaml.
Rebuild the report from completed local outputs with python main.py report. It downloads a checksummed Japanese font into the ignored cache to embed native Japanese correctly. scripts/refresh_analysis.py regenerates figures and the paired BLEU bootstrap. scripts/verify_rerun.py performs GPU checks; scripts/verify_publication.py checks local/remote hashes, current W&B summaries and public raw outputs.
Code map
| Path | Responsibility |
|---|---|
main.py |
CLI and task selection |
src/data.py |
pinned dataset revisions, train-only BPE, grouped splits and token streams |
src/part1/model.py |
decoder, interchangeable FFNs, router and packed-example mask |
src/part2/optimizers.py |
hand-written AdamW, NAdamW, Lion and Muon updates |
src/training.py |
token-budget loop, explicit schedule, eleven evaluations and resume state |
src/evaluation.py |
target-token likelihood, greedy BLEU and source routing |
src/part3/decoding.py |
greedy, beam, top-k and top-p using forward calls and cache |
src/part3/experiment.py |
FP32 ROCStories evaluation and saved continuations |
scripts/collect_routing_controls.py |
per-layer VI/JA and English-target routing |
scripts/build_report.py |
measured tables and page-bounded PDF layout |
Metric definitions and controls
Part 1 PPL scores the complete English target/EOS test stream. Translation BLEU uses the same 2,000 pairs, greedy decoding and 64 new tokens; paired intervals resample triplets, keeping VI/JA together. Block-diagonal masks prevent paired-answer leakage. Part 2 consumes one complete grouped-document training stream and measures token-weighted validation NLL and BLEU over 1,000 cached held-out continuation prefixes at every 0.1 fraction. Part 3 uses the same 1,000 ROCStories prompts for every configuration and 40 new tokens in FP32. Reference PPL is independent of decoding; generated PPL measures model confidence. Aligned accuracy counts missing positions as wrong. All results use seed 42; bootstrap intervals do not quantify training-seed variation.
Public experiments and checkpoints
| Experiment | W&B | Hugging Face |
|---|---|---|
| part1_dense | ji1yoopj | model |
| part1_moe_active | 7pggqpl2 | model |
| part1_moe_shared | b12pwkkc | model |
| part1_moe_top1 | ns6f12bt | model |
| part1_moe_top2 | 66mk8axg | model |
| part2_adamw | kg1mgcbt | model |
| part2_lion | ae1gyqla | model |
| part2_lion_lr1.2e-4 | jlqs7h4h | model |
| part2_lion_lr2e-4 | pks39g67 | model |
| part2_muon | znv1b9rv | model |
| part2_nadamw | mzw7cepr | model |
| part3_beam1 | dky6nu2f | model |
| part3_beam2 | bfhp7276 | model |
| part3_beam4 | vs53f0wm | model |
| part3_greedy | 9qe2l6vg | model |
| part3_topk10_t10 | 4zzrufjg | model |
| part3_topk20_t07 | tldy8xwb | model |
| part3_topk20_t10 | zkdnld9g | model |
| part3_topk50_t10 | n5wqxz9h | model |
| part3_topp080_t10 | 77uzpask | model |
| part3_topp090_t07 | mw221dgz | model |
| part3_topp090_t10 | 7f81nccl | model |
| part3_topp095_t10 | dnn3amki | model |
All eleven trained checkpoint repositories are public, including the two Lion sweep runs. Optimizer repositories retain all ten fraction snapshots each (60 milestone states in total), and final/best/latest resume states and inference exports remain local. All 23 current W&B runs are finished and have evaluation artifacts. Prior full and smoke runs are retained, with tags/notes distinguishing them; no runs were deleted.
Public raw generations and token IDs · W&B project · Course repository · GitHub mirror (private)
Large checkpoints, caches, local metrics, logs and historical result folders are excluded from Git. The personal interview study README stays local. Data manifests and exact pinned revisions are stored in local data/part*/source.json and in the report's references; checkpoint states retain the data fingerprints. Hugging Face stores the large artifacts while Git holds code, report, figures and compact summaries.
Attribution
The implementation used LLM coding assistance. Optimizer mathematics follow Wen et al. and the Muon author reference. No predefined optimizer update or model.generate() implementation is called. The report states the single-seed and limited-tuning constraints.
- Downloads last month
- 239