malaiwah/glm53-fixture-0.1B-fidelity-root-v1

GLM-5.3-Flash-0.1B fixture — root fidelity dataset (hidden form)

The numbers in this dataset are meaningless as model quality. The weights it was captured from are random. inference-optimization/GLM-5.3-Flash-0.1B-A0.1B is an architectural fixture: it has GLM-5.3-Flash's config shape and its exact 154,880-token vocabulary, and none of its training. Nothing here says anything about GLM-5.3-Flash, about quantization quality, or about any model.

What it is for: being a small, complete, runnable example of the three-step fidelity-dataset format — one you can download in a minute and reproduce in seconds, instead of a 31 GB artifact you have to take on faith.

The three steps

step 1  capture   reference weights + panel  ->  fidelity dataset A   (this repo)
step 2  capture   quantized weights + panel  ->  fidelity dataset B
step 3  compare   A, B                       ->  KLD + determinism + a registry receipt
                  A, A'                      ->  reproduction confirmation, exactly 0.0

This repository is A. It is a root: every tensor class is native bf16, the head is native, and scope.policy is native.

What is in it

form	hidden — the lm_head input, not logits
semantic point	after_final_rmsnorm_before_lm_head
records	8 windows
scored positions	16,376 (2,047 per window)
context length	2,048
hidden width	256
vocab size	154,880
capture bytes	8,386,600
the same capture in logit form	10,145,259,520 bytes — 1,210×
head payload	79,298,652 bytes (bf16, shipped so you can replay without the weights)
lane	local-cuda-budget (one NVIDIA L4)

Why hidden form

A hidden-form record costs hidden_size × 2 bytes per scored position; a logit-form record costs vocab_size × 4. Here that is 512 versus 619,520 — 1,210× — and the head is shipped once so anyone can recover the logits exactly:

logits = hidden.float() @ head.float().T

That is the whole replay. The final norm is not applied at replay time: the capture already sits after it.

Panel

The token panel is the final role of brandonmusic/GLM-5.3-Flash-BF16-Teacher-Logits calibration/panel-v1, revision 95f4fdd94bf29989db2e0d1054e4931f55edb6aa — the first 8 of its 25 final windows, to keep the run cheap. The subset is declared in coverage.subset_detail, and panel/panel.json carries the upstream panel.receipt.json digest so the binding can be re-checked against the source. The panel's token ids are valid here because the fixture's vocabulary is byte-for-byte the real model's.

None of the teacher logits from that dataset were used or are needed.

Reproduce it

git clone https://github.com/malaiwah/glm53-flash-fidelity-suite
cd glm53-flash-fidelity-suite
pip install torch transformers safetensors huggingface_hub numpy

# fetch the fixture and the panel (see the commands in the repo's README)
python3 bin/fidelity_dataset.py capture \
    --out ds-root --form hidden --role root --lane local-cuda-budget \
    --engine hf-transformers -- \
    --model inference-optimization/GLM-5.3-Flash-0.1B-A0.1B \
    --model-revision 7c3a6d3dc51732dd8ab230888e06ba8c93a381ac \
    --panel <panel-v1 dir> --panel-role final --windows 8 --device cuda \
    --dataset-id fidelity--glm53fixture0.1B.root.bf16 \
    --dataset-name "GLM-5.3-Flash-0.1B-A0.1B fixture root fidelity dataset (hidden form)"

python3 bin/fidelity_dataset.py verify ds-root

Two cold captures of these weights on one L4 agreed bitwise:

capture_content_digest  d6d16431b88768bc45bafd9464861839c9f4b0114cdd06c0431776e2d8f5961c

and the reproduction confirmation is exact:

python3 bin/fidelity_dataset.py compare --reference ds-root-1 --candidate ds-root-2 \
    --out cmp --self-compare --force-compute
# mean_tokenwise_kld = 0.0 nats, top-1 agreement 1.0

--force-compute runs the full fp64 estimator over all 16,376 × 154,880 logits rather than accepting the capture-digest hash proof. Both agree on exactly 0.0.

Verify what you downloaded

python3 bin/fidelity_dataset.py verify <this directory>

fidelity-dataset.json seals itself; checksums.txt covers every other file; capture.capture_content_digest covers every capture tensor a second time, by content. A flipped byte anywhere is a refusal, not a warning.

What a comparison against it looks like

Against a deliberately crude candidate — round-to-nearest, per-output-row absmax scale, applied to the routed-expert tensors only and to nothing else (bin/toy_quantize.py):

candidate	mean tokenwise KLD (nats)	top-1 agreement
RTN int8, routed experts	0.000154605	99.33%
RTN int4, routed experts	0.004743061	91.58%

These are tooling numbers on random weights. They demonstrate that the comparator produces an ordered, nonzero, reproducible result. They are not a statement about int4 or int8 quantization of anything.

Specification

• Format: docs/FIDELITY-DATASET-SPEC.md
• Card annotation: docs/CARD-ANNOTATION-SPEC.md
• Registry integration: docs/REGISTRY-INTEGRATION.md

Disclosures

• fixture_weights — random weights; every number is tooling evidence only.
• reduced_run_count — one cold capture (run_count: 1; the DET-D4 floor is 5). Cross-run determinism is not asserted by this dataset; it is established by the self-compare above.
• subset_of_panel — 8 of the upstream panel's 25 final windows.

MIT licensed, like the tooling.