bf16_files list | capture_output_dir string | domain_tokens dict | expected_candidate_logit_bytes int64 | input_manifest dict | rows list | schema_version int64 | token_ids dict | total_rows int64 | total_tokens int64 | vocab_size int64 |
|---|---|---|---|---|---|---|---|---|---|---|
[
{
"bytes": 2539102720,
"path": "runs/run-089-bf16-reference-capture-m2-override/bf16-capture-stage/work/rank-0/base-final-quality-logits.bf16",
"rank": 0,
"sha256": "0c1f66c23b18b2c991dd514c758fc9511ba397e4c75ab4ac3fdafcb61afadfa7",
"shape": [
8197,
154880
]
},
{
"bytes":... | /home/js/projects/glm52-shared-h-current/records/kld-plan/capture-output | {
"agentic": 8192,
"code": 8192,
"general": 8192,
"reasoning": 8192
} | 10,150,215,680 | {
"path": "data/manifests/bf16-capture-inputs-v1.json",
"sha256": "6c2464e7d853ed343e3a79c53bc704fae5a70869fd83785834ee1535df9842a3"
} | [
{
"bf16_rank": 2,
"bf16_rank_offset": 0,
"domain": "agentic",
"index": 0,
"row_id": "06c73115b5bf495258a726d3857cfe22c8027b830685a50e58530c496f8f4477",
"token_count": 41,
"token_ids_sha256": "0872ab9d0cf86b15faf29c66add8966b170f10d91e47ebccd47be503277c3406",
"token_offset": 0
},
... | 1 | {
"bytes": 131072,
"dtype": "uint32",
"endianness": "little",
"path": "data/calibration/bf16-capture-inputs-v1/final-quality-token-ids.i32",
"sha256": "81c1f40598e1f8085e1a3b86f1176ee0830716065ac35534f75fab93c44bc592"
} | 103 | 32,768 | 154,880 |
GLM-5.2 KLD item13 Calibration Panel — Plan & Row Hashes
Reproduction package for the calibration-domain KLD measurements cited on jpsequeira/GLM-5.2-EXL3-TR3-3.40bpw-KVarN-K4V2 (the "0.0558 of-record" / "0.0555 on tr3.2" panel).
Read the model card's retraction section first. These numbers are calibration-domain (in-distribution): the corpus family is the same one the quantization allocation was tuned on. They measure in-distribution serving fidelity, not held-out generalization. Held-out numbers (wikitext-2, festr2 reference) are in the card's correction section.
Contents
capture-plan.json— the sealed 103-row plan (SHA256072e37fabc39c616320ada867c48741ab3b9f255255ac7b37b2f0aeedace4e4e). Rows carrytoken_ids_sha256= SHA256 of the little-endian int32 token-id array (theitem13_prompt_logits.pyconvention, lines 241–266).rows.tsv— flat index / row_id / domain / token_count / token_ids_sha256 listing for quick verification.
Panel composition
103 rows, 32,768 tokens total: agentic 27 rows / code 14 / general 46 / reasoning 16 — 8,192 tokens per domain.
Reproducing
Serve the checkpoint (image jmpsequeira/glm52-kvarn-k4v2:tr3.2,
kv_cache_dtype kvarn_mla_k2_g64), replay the plan rows by raw token ids
with the capture hook (see the model card quick start for the image; the
client contract is documented in the plan itself), and compare the captured
logits against a BF16 reference of your own capture under identical
conditions. Domain split and evaluator keys are in the plan.
Reference values measured with this plan: forward KLD 0.0558 (dtype-cutover panel), 0.0555 (tr3.2, APC on).
- Downloads last month
- 37