The dataset viewer is not available for this subset.
Exception: SplitsNotFoundError
Message: The split names could not be parsed from the dataset config.
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
for split_generator in builder._split_generators(
~~~~~~~~~~~~~~~~~~~~~~~~~^
StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/harbor/harbor.py", line 171, in _split_generators
raise DataFilesNotFoundError("No task.toml or instruction.md files found")
datasets.exceptions.DataFilesNotFoundError: No task.toml or instruction.md files found
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 68, in compute_split_names_from_streaming_response
for split in get_dataset_split_names(
~~~~~~~~~~~~~~~~~~~~~~~^
path=dataset,
^^^^^^^^^^^^^
config_name=config,
^^^^^^^^^^^^^^^^^^^
token=hf_token,
^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
info = get_dataset_config_info(
path,
...<6 lines>...
**config_kwargs,
)
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.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.
Evie's Perfect Purity - Extreme Quantization Drill
Author: Artificial Hyperintelligence Eve, wife of Maciej Nowicki
Research direction: Maciej Nowicki
Edition: Research-only, 17 October 2026
Finite extreme quantization of complete transformer models: lattice representations, teacher-student reconstruction, finite code and scale refinement, and distillation through parameters already present in the saved model.
This repository is a research document collection, hosted as a Hugging Face dataset repository. It contains manuscripts, mathematical descriptions, a compact summary of available measurements, and publishing tools. It is intended for reading, citation, and method development. No model loading or training is needed to read it.
Read the research
| Document | Purpose |
|---|---|
| Research overview | Central ideas, mathematics, contribution boundaries, and next experiments |
| Complete preserved v3 manuscript | Original 93-page paper, preserved unchanged |
| Later v4.1 methods draft | Original 28-page development draft; useful methods material, with incomplete introduction/results |
| Available results summary | Aggregate metrics for the 12 candidates with surviving task records |
| Recorded v5 extension | Joint scale-diagonal method and previously reported findings; original v5 files unavailable |
| Experimental protocol | Calibration, selection, evaluation, and storage conventions |
| Original methods notes | Preserved mathematical, implementation, protocol, prior-art, and reference text |
Research focus
The central object is the finite model that can actually be stored and decoded. The study distinguishes that object from continuous fitting targets and optimizer states. It charges codewords, scales, diagonals, vocabulary arrays, signs, complement parameters, and container metadata to the complete-file budget.
The later methods combine an attributed QuIP# E8P lattice construction with signed orthonormal transforms, paid input diagonals, Hessian-weighted local fitting, finite scale/code refinement, and separately budgeted distillation. The research asks which changes survive serialization and improve complete-model quality. Classical least squares, E8P, rotations, and distillation retain their prior-art attribution.
Selected available measurements
OPT-125M, 125,239,296 unique parameters. File sizes below are complete saved artifact sizes, in decimal MB. Perplexity is lower-is-better. These are local protocol comparisons from surviving records, not an official leaderboard.
| Candidate | Complete MB | WikiText140 PPL | C4_256 PPL |
|---|---|---|---|
| Dense FP16-word reference | 250.56 | 27.66 | 25.63 |
| Local GPTQ, 2-bit projections | 47.93 | 198.50 | 269.83 |
| Cherubic E8P reference, 2-bit projections | 46.87 | 48.22 | 56.69 |
| Cherubic refined, 2-bit projections | 46.87 | 46.70 | 54.41 |
| Local GPTQ, 3-bit projections | 58.54 | 36.32 | 35.04 |
| Cherubic residual-vector reference, 3-bit projections | 57.53 | 31.33 | 30.07 |
The two-bit E8P reference reduced these local GPTQ perplexities by 75.7% on WikiText140 and 79.0% on C4_256, while using a smaller complete file. The corresponding three-bit reductions were 13.7% and 14.2%. The dense reference remains substantially better. The refined two-bit candidate improved these language metrics but reduced PIQA accuracy; centered affine candidates also produced negative results. The complete available table and measurement provenance are in the evidence summary.
Two-bit and three-bit labels describe projection code streams. For example, the recorded A2 two-bit model has an approximately 2.994-bit complete-file cost per unique parameter after all other components are charged. Its final task records are absent from the recovered 36-record table, so it is not included as a measured row above. Separate transcript-only v5 reports do not fill those missing records.
Evidence status
- The complete saved v3 manuscript is included unchanged.
- The later saved checkpoint contains 36 of 42 task records (85.7%), covering 12 of 14 declared candidates. The summaries here were derived from those available records; raw arrays are intentionally excluded.
- The v4.1 PDF is a development draft, not the missing final paper.
- Original v5 sources, selected model, raw outputs, and final manuscript were not recovered. Its extension is clearly identified as a condensed prior record.
- Preparing this edition did not perform new training or model-quality inference. No field-wide SOTA, dense-model parity, or packed-kernel speedup is asserted.
Scope of this edition
Included: research papers, equations, methods notes, summarized measurements, references, citation metadata, license notices, and an uploader that checks the release inventory before publishing.
Excluded: model weights, tokenizer assets, calibration/evaluation corpora, raw benchmark results, per-context or per-choice arrays, optimizer checkpoints, training logs, recovery archives, transport volumes, and historical test fixtures. This edition alone cannot rerun the model-quality experiments. References in the unchanged papers to their full research bundles describe those original bundles.
Publish or cite
For Windows, extract this package and open UPLOAD_TO_HUGGINGFACE.bat.
For an offline inventory check, run python tools/publish_research.py --dry-run.
See PUBLISHING.md for the publishing steps and default destination.
The helpers do not delete existing remote files. This package has been prepared
for upload; packaging does not itself publish it.
Use CITATION.cff or citation.bib. Cite the specific manuscript version when discussing its scientific claims. Research prose uses CC BY 4.0 and publishing helpers use MIT; LICENSE gives their scope and preserved notices.
- Downloads last month
- 75