Dataset Viewer
The dataset viewer is not available for this split.
Cannot extract the features (columns) for the split 'train' of the config 'default' of the dataset.
Error code: FeaturesError
Exception: ArrowInvalid
Message: Schema at index 1 was different:
tag: string
name: string
levels: string
K: int64
d: int64
ckpt: string
best_val: double
test_acc: double
logged_test_acc: double
vs
tag: string
name: string
K: int64
d: int64
n_active: int64
usage_frac: double
eff_perplexity: double
traj_div: double
q_error: double
recon_err: double
n_collapsed_dims: int64
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/split/first_rows.py", line 247, in compute_first_rows_from_streaming_response
iterable_dataset = iterable_dataset._resolve_features()
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 4376, in _resolve_features
features = _infer_features_from_batch(self.with_format(None)._head())
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 2658, in _head
return next(iter(self.iter(batch_size=n)))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 2836, in iter
for key, pa_table in ex_iterable.iter_arrow():
^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 2374, in _iter_arrow
yield from self.ex_iterable._iter_arrow()
File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 580, in _iter_arrow
yield new_key, pa.Table.from_batches(chunks_buffer)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "pyarrow/table.pxi", line 5039, in pyarrow.lib.Table.from_batches
File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
pyarrow.lib.ArrowInvalid: Schema at index 1 was different:
tag: string
name: string
levels: string
K: int64
d: int64
ckpt: string
best_val: double
test_acc: double
logged_test_acc: double
vs
tag: string
name: string
K: int64
d: int64
n_active: int64
usage_frac: double
eff_perplexity: double
traj_div: double
q_error: double
recon_err: double
n_collapsed_dims: int64Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
COCONUT-FSQ bottleneck ablation — cached artifacts
Companion caches for Part I of pau_amargantialvarez_408221.ipynb (Pau Amargant Alvarez, MNLP).
FSQ-COCONUT fine-tuned from plain COCONUT (gsm-coconut/checkpoint_25), varying only fsq_levels.
Each run taken at its best-validation checkpoint and evaluated on the GSM8K test split (N=1319),
through the same harness as the interpretability part.
| config | fsq_levels | K | d | best val | test acc | active codes | eff. ppl | collapsed dims |
|---|---|---|---|---|---|---|---|---|
| FSQ K=1000 (orig) | [8,5,5,5] | 1000 | 4 | — | 0.248 | 849 | 507 | 0/4 |
| dim6 | [4,4,4,4,4,4] | 4096 | 6 | 0.274 | 0.240 | 2097 | 975 | 0/6 |
| dim4 | [8,8,8,8] | 4096 | 4 | 0.282 | 0.262 | 2197 | 1324 | 0/4 |
| k65k | [7,5,5,5,5,5,3] | 65625 | 7 | 0.304 | 0.288 | 4280 | 2681 | 0/7 |
Plain COCONUT baseline (continuous latents): 0.344. All FSQ variants underperform it; the largest codebook (k65k) is best and no codebook collapses — the bottleneck is the encoder squeeze, not the grid.
Files
latents_fsq_{dim4,dim6,k65k}.pt— per-slot recomputed dumps: pre_vq/z/quant/recon (N,6,·) + code_id (N,6), with recomputed exact-match.fsq_ablation.csv— accuracy table (best-val checkpoint, recomputed + logged test acc).fsq_codeusage.csv— active codes, effective perplexity, trajectory diversity, q-error, recon-error, collapsed dims.fsq_zstd.csv— per-dimension encoder z_std (collapse check).fsq_ablation_summary.json— machine-readable summary.
- Downloads last month
- 51