checks_run list | error_count int64 | errors list | format_version int64 | receipt_sha256 string | schema string | structural_status string | subject string | warning_count int64 | warnings list |
|---|---|---|---|---|---|---|---|---|---|
[
"pre-seal"
] | 0 | [] | 1 | 36e448ea0fa4fa93e2976e07fce308c04c918cc31a5786d7347bcee25dd161e1 | malaiwah.fidelity-structural-validation.v1 | sealed | dataset:fidelity--571d4d7defe5266c928e7dcbf77c2791 | 0 | [] |
HF workflow 571d4d7defe5266c928e7dcbf77c2791
A root fidelity dataset in hidden form, produced by engines/tools/hf_capture.py from malaiwah/glm-moe-dsa-tiny-random-bf16.
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 engines/tools/hidden_replay.py and as Festr's kimi-k3 hidden-replay qualification.
What is here
| records | 4 |
| scored positions | 252 |
| context length | 64 |
| hidden width | 64 |
| vocab size | 260 |
| capture bytes (hidden form) | 33404 |
| the same capture in logit form would be | 262080 bytes (8x) |
| scope policy | native |
| lane | other |
Does the model still generate sensibly
The capital of France is -> � (p=0.0067, entropy 5.548 of a possible 5.561 nats). Recorded, not enforced. One extra window through the same schedule; its hidden state was discarded and is not part of this dataset.
License
This dataset redistributes the captured model's native output-head weights. The source license and required notice are reproduced byte-for-byte in LICENSE.
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
- Downloads last month
- 9