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
model: string
calib_seed: int64
seqlen: int64
ckpt: string
prefix: string
arm: string
eval: string
nlls: list<item: double>
child 0, item: double
nll: double
obj_fresh_before: double
g_tokens: int64
obj_ratio: double
nll_wt2: double
flips: struct<q_proj: double, depth0: double, k_proj: double, v_proj: double, o_proj: double, gate_proj: do (... 89 chars omitted)
child 0, q_proj: double
child 1, depth0: double
child 2, k_proj: double
child 3, v_proj: double
child 4, o_proj: double
child 5, gate_proj: double
child 6, up_proj: double
child 7, down_proj: double
child 8, depth1: double
child 9, depth2: double
child 10, depth3: double
peak_vram_mib: double
calib_source: string
scope_s: double
bpw: double
kl_fp16: double
nll_c4: double
obj_fresh: double
quant_s: double
opt_s: double
obj_fresh_after: double
n_probe: int64
to
{'model': Value('string'), 'calib_seed': Value('int64'), 'seqlen': Value('int64'), 'calib_source': Value('string'), 'n_probe': Value('int64'), 'g_tokens': Value('int64'), 'arm': Value('string'), 'eval': Value('string'), 'nlls': List(Value('float64')), 'nll': Value('float64'), 'bpw': Value('float64'), 'flips': {'q_proj': Value('float64'), 'depth0': Value('float64'), 'k_proj': Value('float64'), 'v_proj': Value('float64'), 'o_proj': Value('float64'), 'gate_proj': Value('float64'), 'up_proj': Value('float64'), 'down_proj': Value('float64'), 'depth1': Value('float64'), 'depth2': Value('float64'), 'depth3': Value('float64')}, 'obj_ratio': Value('float64'), 'obj_fresh': Value('float64'), 'obj_fresh_before': Value('float64'), 'obj_fresh_after': Value('float64'), 'quant_s': Value('float64'), 'opt_s': Value('float64'), 'scope_s': Value('float64'), 'peak_vram_mib': Value('float64'), 'nll_wt2': Value('float64'), 'nll_c4': Value('float64'), 'kl_fp16': Value('float64')}
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
calib_seed: int64
seqlen: int64
ckpt: string
prefix: string
arm: string
eval: string
nlls: list<item: double>
child 0, item: double
nll: double
obj_fresh_before: double
g_tokens: int64
obj_ratio: double
nll_wt2: double
flips: struct<q_proj: double, depth0: double, k_proj: double, v_proj: double, o_proj: double, gate_proj: do (... 89 chars omitted)
child 0, q_proj: double
child 1, depth0: double
child 2, k_proj: double
child 3, v_proj: double
child 4, o_proj: double
child 5, gate_proj: double
child 6, up_proj: double
child 7, down_proj: double
child 8, depth1: double
child 9, depth2: double
child 10, depth3: double
peak_vram_mib: double
calib_source: string
scope_s: double
bpw: double
kl_fp16: double
nll_c4: double
obj_fresh: double
quant_s: double
opt_s: double
obj_fresh_after: double
n_probe: int64
to
{'model': Value('string'), 'calib_seed': Value('int64'), 'seqlen': Value('int64'), 'calib_source': Value('string'), 'n_probe': Value('int64'), 'g_tokens': Value('int64'), 'arm': Value('string'), 'eval': Value('string'), 'nlls': List(Value('float64')), 'nll': Value('float64'), 'bpw': Value('float64'), 'flips': {'q_proj': Value('float64'), 'depth0': Value('float64'), 'k_proj': Value('float64'), 'v_proj': Value('float64'), 'o_proj': Value('float64'), 'gate_proj': Value('float64'), 'up_proj': Value('float64'), 'down_proj': Value('float64'), 'depth1': Value('float64'), 'depth2': Value('float64'), 'depth3': Value('float64')}, 'obj_ratio': Value('float64'), 'obj_fresh': Value('float64'), 'obj_fresh_before': Value('float64'), 'obj_fresh_after': Value('float64'), 'quant_s': Value('float64'), 'opt_s': Value('float64'), 'scope_s': Value('float64'), 'peak_vram_mib': Value('float64'), 'nll_wt2': Value('float64'), 'nll_c4': Value('float64'), 'kl_fp16': Value('float64')}
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.
llm-weight-compression: evidence
Per-window evaluation data and raw diagnostics behind every number in rosecky/llm-weight-compression-public: a pre-registered study of refining GPTQ's integer codes against the output of the enclosing attention/MLP module instead of the single layer, on Qwen2.5-0.5B, Qwen3-0.6B, Llama-3.2-1B and Qwen2.5-1.5B at exactly 3.25 bits per weight, plus rate–distortion, estimation-budget and 2-bit vector-quantized points on Qwen2.5-0.5B.
The long-form write-up is in ARTICLE.md in this repo.
Layout
validation/ frozen-protocol runs: one .jsonl per (model, calibration draw) with
per-window mean NLL for every arm on the full wikitext-2 test (146 windows
of 2048 tokens) and 512 fixed C4 validation windows, plus a summary row per
arm (bpw, KL to fp16, objective ratios, flips, timings, peak VRAM); the
matching .log is the run's stdout
diagnostics/ exploratory-phase runs (objective ladder, superadditivity, VQ factorial),
probe-budget rescoring, rotation landscape, rate-allocation oracle, oracle
prediction features
box_scripts/ the exact shell scripts run on the rented GPU instance
README.md file-by-file map and the pre-registration commit for each stage
Each validation/*.jsonl row is either {"arm", "eval": "wt2"|"c4", "nlls": [...], "nll": mean} or a per-arm summary with bpw, nll_wt2, nll_c4, kl_fp16, obj_ratio,
obj_fresh, flips, opt_s, peak_vram_mib. Windows are in corpus order, so any two
arms of the same file are paired window by window.
Regenerate the tables
git clone https://github.com/rosecky/llm-weight-compression-public
python scripts/analyze_valid.py --print # four-model table, paired bootstraps within draw
python scripts/analyze_rd.py # rate–distortion, equivalent bits
python scripts/analyze_d3.py # halves, calibration domain, estimation budgets
with the analysis scripts pointed at this dataset's validation/ directory. No GPU is
needed.
Provenance
Method and protocol were frozen and committed before any validation run
(docs/FROZEN_PROTOCOL.md in the repository); diagnostic addenda were pre-registered with
interpretation tables before running. The 2-bit checkpoints evaluated in
validation/p4_* and validation/d4_* were produced by a per-block VQ quantiser from the
embedding-quantization project (d=4, K=256, ~2.1 bpw, input-side rotation) that is not part
of this release; the dequantised evaluations are.
Author: Jan Rosecký, September 2026. Experiments run with Claude (Anthropic) as an autonomous research agent under the author's direction. License: MIT.
- Downloads last month
- 23