Datasetpublic

malaiwah/glm52-fidelity-root-v1

fidelity--glm52.malaiwah.root.bf16 A root fidelity dataset in hidden form, produced by engines/tools/hf_capture.py from zai-org/GLM-5.2. The cut the final hidden state handed to lm_head -- after the text model's final norm and immediately before the head matmul -- captured as the head module's input via torch.nn.Module.register_forward_pre_hook; replay applies the head ONLY (no final norm at replay time: the capture already sits after it). Same cut as… See the full description on the dataset page: https://huggingface.co/datasets/malaiwah/glm52-fidelity-root-v1.

sourceHugging Facemitupdated 21d agoView on Hugging Face

0likes116downloads

Dataset Card

fidelity--glm52.malaiwah.root.bf16

A root fidelity dataset in hidden form, produced by engines/tools/hf_capture.py from zai-org/GLM-5.2.

The cut

the final hidden state handed to lmhead -- after the text model's final norm and immediately before the head matmul -- captured as the head module's input via torch.nn.Module.registerforwardprehook; replay applies the head ONLY (no final norm at replay time: the capture already sits after it). Same cut as engines/tools/hidden_replay.py and as Festr's kimi-k3 hidden-replay qualification.

What is here


records	25
scored positions	51175
context length	2048
hidden width	6144
vocab size	154880
capture bytes (hidden form)	628845100
the same capture in logit form would be	31703936000 bytes (50x)
scope policy	native
lane	streaming

Does the model still generate sensibly

The capital of France is -> Paris (p=0.4886, entropy 3.083 of a possible 11.950 nats). Enforced against 'Paris'. One extra window through the same schedule; its hidden state was discarded and is not part of this dataset.

How to check it

bin/fidelity-dataset verify <this directory>

Full specification: <https://github.com/malaiwah/quant-fidelity-suite/blob/main/docs/FIDELITY-DATASET-SPEC.md>