malaiwah/fruit-fidelity-root-v1

GLM-5.2-SIQ-Fruit — root fidelity dataset (hidden form)

This is the reference yardstick for the GLM-5.2-SIQ-Fruit family: one bf16 forward pass of malaiwah/GLM-5.2-SIQ-Fruit-bf16 over a sealed 16-window token panel, captured at the lm_head input and sealed so that anybody can compare a quantized capture against it without the weights, without our infrastructure, and without re-running the reference.

Fruit is a 5.04B-parameter / 0.46B-active GLM-5.2-architecture serving proxy that the publisher of this dataset trained himself. That matters here for one reason: the reference is unambiguous. There is no borrowed teacher, no third-party checkpoint whose provenance has to be argued about, and no lane bridged to another lane. The measurement is his model against his own bf16.

Scope of the claim. These numbers describe distribution agreement on this panel. They are not a quality benchmark, and Fruit is explicitly a CI fixture and kernel-development vehicle rather than an assistant.

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, scope.policy is native, and lossy_codec is null.

What is in it

form	hidden — the lm_head input, not logits
semantic point	after_final_rmsnorm_before_lm_head
records	16 windows
scored positions	32,752 (2,047 per window)
context length	2,048
hidden width	1,024
vocab size	154,880
capture bytes	67,080,288
the same capture in logit form	20,290,519,040 bytes — 302×
head payload	317,194,334 bytes (bf16, shipped so you can replay without the weights)
lane	local-cuda-budget (one NVIDIA L4, spot)

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 2,048 versus 619,520 — 302× — 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.

The panel is Fruit's own

panel--fruit.malaiwah.heldout-v1, built for this model and for nothing else.

Fruit shares its 154,880-token vocabulary with GLM-5.3-Flash, so the sealed panels published for that model are numerically valid against Fruit. They are deliberately not used here. A panel selected for another model invites its numbers to be ranked against that model's numbers, which is exactly the cross-model comparison this registry's rules forbid. Fruit gets its own panel id so a Fruit number can never be silently ranked against a GLM-5.3-Flash one.

panel id	panel--fruit.malaiwah.heldout-v1
windows	16 (8 literary + 8 scientific)
context length	2,048, every causal prediction position scored, nothing dropped
suite_token_hash_sha256	a6d367cc3ba448800372dee435d2bb4f536d23ca68843628832fa3b122ceabe1
corpus	malaiwah/qwen38-27b-fidelity-suite-v5 @ 7797fcce3ffed62b99871348887f4626dc9b2b3b, corpus/text/

How it was built (no RNG anywhere — panel/panel-receipt.json carries the rule, every source document's sha256, and the exact token slice used):

1. strata sorted ascending; within each stratum, documents sorted by file name;
2. each document tokenized whole with add_special_tokens=False;
3. eligible only if it yields ≥ 4,096 tokens; its window is tokens[2048:4096] — the leading 2,048 tokens are dropped because the head of a real document is title pages and boilerplate, the least representative text in it;
4. the first 8 eligible documents per stratum are taken.

python3 k6/tools/build_token_panel.py \
    --corpus-dir <corpus/text> --stratum literary --stratum scientific \
    --corpus-repository datasets/malaiwah/qwen38-27b-fidelity-suite-v5 \
    --corpus-revision 7797fcce3ffed62b99871348887f4626dc9b2b3b \
    --corpus-path-prefix corpus/text/ \
    --tokenizer <a local copy of the Fruit tokenizer> \
    --tokenizer-repository malaiwah/GLM-5.2-SIQ-Fruit-bf16 \
    --tokenizer-revision ef68013aa6e16453cf52b5b77647f72fbe258c3c \
    --tokenizer-id glm-5.2-siq-fruit \
    --windows-per-stratum 8 --context-length 2048 --skip-tokens 2048 \
    --panel-id panel--fruit.malaiwah.heldout-v1 --out <panel dir>

Separation, stated honestly. Fruit's published pretraining recipe names nine sources: FineWeb-Edu, English and Chinese Wikipedia, TinyStories, two GLM-5.2 distillation corpora, REAP calibration text, SPDX licence text, and code. This panel draws only from the literary (Project Gutenberg) and scientific (arXiv titles and abstracts) strata, neither of which appears in that list. That is source-level separation. No lexical or n-gram scan against Fruit's actual training shards was run, so incidental overlap through a web-crawl source such as FineWeb-Edu is not excluded. panel.contamination.checked is false and says so.

What this lane actually ran

Stock transformers 5.16.1 implements glm_moe_dsa natively (no trust_remote_code), and loads 4,572,134,656 of Fruit's 5.04B parameters. It reports the DSA lightning-indexer tensors for layers 3–13 and the entire MTP draft layer 13 as unexpected keys and drops them, so the forward pass is dense MLA attention with no speculative decoding. That is a property of the lane, not of the checkpoint, it is what the model card documents for this path, and it is identical for every capture compared against this one — which is what keeps the comparison honest.

Determinism

Two cold captures, two separate processes on the same L4, agreed bitwise:

capture_content_digest  b417acc22b8aa7f3294b8e62c4b619bc5051aef9fd8a073602572a30af6b3e1c

The two runs' dataset_sha256 values differ, because a manifest embeds timestamps and a cold-run label. That is exactly why determinism evidence is taken over tensor content and never over a container digest.

Reproduce it

git clone https://github.com/malaiwah/glm53-flash-fidelity-suite
cd glm53-flash-fidelity-suite
pip install torch "transformers==5.16.1" safetensors huggingface_hub numpy

python3 bin/fidelity_dataset.py capture \
    --out ds-root --form hidden --role root --lane local-cuda-budget \
    --engine hf-transformers -- \
    --model malaiwah/GLM-5.2-SIQ-Fruit-bf16 \
    --model-revision ef68013aa6e16453cf52b5b77647f72fbe258c3c \
    --weights-repository malaiwah/GLM-5.2-SIQ-Fruit-bf16 \
    --panel <panel dir> --panel-role final --device cuda \
    --panel-id panel--fruit.malaiwah.heldout-v1 \
    --tokenizer-id glm-5.2-siq-fruit \
    --dataset-id fidelity--fruit.malaiwah.root.bf16 \
    --dataset-name "GLM-5.2-SIQ-Fruit bf16 root fidelity dataset (hidden form)"

python3 bin/fidelity_dataset.py verify ds-root

The whole capture takes about 18 seconds of GPU time.

Compare something against it

python3 bin/fidelity_dataset.py compare \
    --reference hf://malaiwah/fruit-fidelity-root-v1 \
    --candidate <your quant capture> \
    --out cmp --verify-tensors

and the reproduction confirmation, which must be exactly zero:

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

--force-compute runs the full fp64 estimator over all 32,752 × 154,880 logits rather than accepting the capture-digest hash proof.

Because a candidate captured with the same engine on the same lane shares this capture's numerical environment, the comparison floor is structurally zero rather than something subtracted. That is the whole point of separating capture from comparison, and it is why a same-lane number here is not comparable to a cross-stack number measured elsewhere.

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; panel/panel-receipt.json ships verbatim so the panel's own digest has a preimage you can check. A flipped byte anywhere is a refusal, not a warning.

Disclosures

• reduced_run_count — this published tree is one cold capture (run_count: 1). Cross-run determinism is not asserted by the tree itself; it is established by the two-cold-run self-compare above.
• weak_contamination_guard — source-level separation only; no n-gram scan against Fruit's training shards. See the panel section.
• architecture_subset_loaded — stock transformers drops Fruit's DSA indexer (layers 3–13) and MTP layer 13. Every capture on this lane drops the same set.

Specification

• Format: docs/FIDELITY-DATASET-SPEC.md
• Card annotation: docs/CARD-ANNOTATION-SPEC.md
• Registry: malaiwah/quant-fidelity-registry

MIT licensed, like the tooling.