Datasetpublic

anonymousmaharaj/GLM-5.3-Flash-Research-Artifacts

GLM-5.3 Flash research artifacts This is a research dataset from Alexey's attempt to make GLM-5.3 Flash useful on 4x RTX 3090. It contains numerical measurements, synthetic reference tests and a license-reviewed subset of teacher distributions captured on RTX 6000 PRO hardware. The experiment did not produce a quality-qualified production model. The teacher was RedHatAI/GLM-5.3-Flash-NVFP4, revision 36c184c6cda000a481711306df5adde42f63321a, based on Z.AI's GLM-5.3 Flash. These… See the full description on the dataset page: https://huggingface.co/datasets/anonymousmaharaj/GLM-5.3-Flash-Research-Artifacts.

sourceHugging Faceotherupdated 20d agoView on Hugging Face

0likes69downloads

Dataset Card

GLM-5.3 Flash research artifacts

This is a research dataset from Alexey's attempt to make GLM-5.3 Flash useful on 4x RTX 3090. It contains numerical measurements, synthetic reference tests and a license-reviewed subset of teacher distributions captured on RTX 6000 PRO hardware. The experiment did not produce a quality-qualified production model.

The teacher was RedHatAI/GLM-5.3-Flash-NVFP4, revision 36c184c6cda000a481711306df5adde42f63321a, based on Z.AI's GLM-5.3 Flash. These measurements describe that NVFP4 checkpoint. They do not measure its loss relative to BF16.

Files

The public package separates numerical summaries from data that contains source text.

Directory	Contents
`reap-atlas/`	REAP importance, ranking stability, activation statistics, router dynamics, quantization statistics and small paired-Vision/pruning diagnostics. JSON summaries only.
`synthetic-reference/`	52 deterministic synthetic tasks with requests, programmatically drawn images, responses, generated-token top-64 distributions and receipts. Also 2 repeatability runs of 16 requests each.
`teacher-glaive/`	599 Glaive-derived samples with prompt token IDs, teacher top-64 log probabilities, tail mass and validity masks. 314,697 positions, of which 314,098 are valid loss positions.
`negative-results/`	Numerical pilot reports for Q3/Q4 expert quantization, router controls and block LoRA. No adapter or checkpoint weights.
`runtime-measurements/`	Raw measurements from the experimental 3090 runtime. Individual configurations and acceptance limits remain part of the evidence.
`producers/`	Capture source code and a pinned dataset source map. Historical scripts require their documented runtime and paths; importing or inspecting the data needs no GPU.

manifest.json lists public files, sizes and SHA-256 hashes. file-provenance.json identifies copied artifacts. Files in a synthetic test directory retain their original receipts. Original paths inside those artifacts describe the collection machine and are not expected to exist on your machine.

Teacher array alignment

In teacher-glaive, row t predicts the input token at position t from the preceding tokens. loss_mask excludes the first position. Each sample contains input.json, teacher.npz, the original receipt.json and a redistribution notice. The NPZ fields are token_ids, top_ids, top_logprobs, actual_token_logprob, tail_mass and loss_mask.

The synthetic reference NPZ files instead record distributions for generated tokens conditioned on the multimodal prompt and preceding generated tokens. They include prompt_token_ids and generated_token_ids. These 2 formats must not be given the same training alignment.

The saved top-64 values do not contain the full vocabulary distribution. Tail mass records the probability outside that set. Load NPZ files with numpy.load(path, allow_pickle=False) and obtain the matching tokenizer from the pinned source model.

Interpretation

All 52 expanded synthetic tasks passed their structured checks. The repeatability runs recorded 15/16 and 16/16 passes. These are narrow diagnostics with published answers; they are unsuitable as a secret holdout and do not establish general coding, agent or Vision quality. The collection used eager execution and is not a production throughput benchmark.

The mixed source archive contained 7,934 samples and 5,001,903 positions. This release contains only its 599 Glaive samples. Do not describe it as a public 5M-token teacher dataset. corpus-source-counts.json documents the original composition.

We have withheld the remaining raw prompts, their recoverable token IDs, associated teacher arrays and dense per-sample activations pending source-level redistribution checks. Aggregate measurements remain useful without reproducing those inputs. License policy explains each decision and links to pinned upstream sources.

Reuse and attribution

Alexey's independent measurements, synthetic tasks, generated reference artifacts and code are released under MIT, to the extent Alexey holds rights in them. The Glaive-derived subset retains Apache-2.0. This repository does not relicense external datasets or model weights. See NOTICE.md and the full texts in LICENSES/.

Please identify the checkpoint and capture configuration when reusing a result, preserve negative results and report which subset you used. Neither Z.AI, Red Hat nor the dataset authors endorse these experiments.