Datasets:
Formats:
csv
Languages:
English
Size:
< 1K
Tags:
research
post-training-quantization
llm-compression
extreme-quantization
finite-codec
low-rank
License:
| # Complete paid v1.0 selection | |
| The final release selector retains sixty decoded reserve checkpoints and | |
| four paid zero-reserve controls. Each sparse configuration uses one of three | |
| base families (hybrid fork, binary, or scalar GPTQ2), columns16 or256, and | |
| epochs0–4. Each matched dense-factor configuration uses the same three bases, | |
| ranks3 or51, and epochs0–4. Bare hybrid, binary, scalar GPTQ2 and scalar GPTQ1 | |
| controls remain eligible. Every cost is the actual complete EEB1 byte count. | |
| The numerical selector source and decision rule are frozen before final | |
| catalog selection. All sixty checkpoints remain visible, including those | |
| excluded by the training incumbent guard. A checkpoint is admitted exactly | |
| when its decoded native training teacher-category KL is no greater than the | |
| actual epoch0 KL in the same configuration. The selector recomputes that | |
| predicate from retained scores and rejects inconsistent flags or incumbent | |
| values. The four bare controls are always admitted. | |
| The primary cap is the complete bare scalar GPTQ2 bundle size. The extreme | |
| cap is the complete bare scalar GPTQ1 bundle size. Within each cap, the | |
| selector minimizes development selected32-plus-direct-tail teacher KL across | |
| all admitted checkpoints and all bare controls. It uses the unrounded tuple | |
| `(development_kl, complete_block_payload_bytes, name)` to break ties. A | |
| checkpoint that wins its own configuration but exceeds a cap cannot prevent | |
| a cheaper snapshot of that configuration from entering global selection. | |
| The same selection is also reported with the base restricted to hybrid fork, | |
| under both caps. These source-path variants remain visible when a binary, | |
| scalar or dense-factor control wins the global comparison. A missing eligible | |
| source-path variant is reported by omitting that role, without inventing an | |
| in-cap candidate. No role is chosen from fresh next-token loss. | |
| ## Native profile catalog | |
| The diagnostic retains all twelve epoch0 initializer labels, all twelve | |
| unconstrained admitted development winners within their configurations, all | |
| four bare controls, and the selected role aliases. The selected roles may | |
| introduce additional states if their constrained optima differ from these | |
| configuration winners. Exact native state hashes include all six installed | |
| FP32 weights and installed fc2 bias. Labels with identical states reuse one | |
| native quality score after byte-identical installation verification; their | |
| labels do not add independent evidence. | |
| ## Evidence order and commands | |
| From the release directory: | |
| ```bash | |
| python v1_select_release.py --stage freeze | |
| python v1_select_release.py --stage prepare-gptq1 | |
| python v1_select_release.py --stage score-gptq1 --frozen-protocol-sha256 EXACT_GPTQ1_PROTOCOL_SHA256 | |
| python v1_select_release.py --stage propose | |
| ``` | |
| The fourth bare scalar GPTQ1 control has its own paid bundle, numerical | |
| source/protocol freeze, calibration-only start journal and receipt. It is | |
| scored on the same64 calibration contexts and fixed teacher selected IDs as | |
| the three other controls. This control stage uses no labels, optimizer, | |
| fresh tokens or next-token loss. | |
| `results/v1_selection_proposal.json` commits the entire64-record catalog, | |
| actual budget anchors, eligible sets, selected tuple for each role, paid | |
| native profile catalog, source hashes and fitting evidence hashes. An | |
| independent auditor then parses the paid states and replays the selected | |
| native calibration objectives. Its receipt commits the proposal hash. | |
| ```bash | |
| python v1_select_release.py --stage finalize --independent-evidence results/v1_calibration_objective_evidence.json | |
| ``` | |
| This produces immutable `results/v1_model_selection.json`, which binds both | |
| the proposal and independent receipt. The proposal never includes the later | |
| receipt hash, avoiding a circular hash dependency. The native diagnostic | |
| then freezes this final selection, all bundle values, model state and fresh | |
| tokens before fresh scoring. Fitting and selection receipts remain separate | |
| from the fresh native loss journal. | |
| ## Interpretation | |
| The development portion is the last16 of the same64 contexts used for base | |
| quantizer fitting; it is a reserve-selection development split, not an | |
| independent quantizer holdout. Both paid caps are control-specific complete | |
| wire budgets. A development-selected winner is not a proof of optimality | |
| outside the retained finite family, and one first-block native diagnostic | |
| cannot establish full-model or industry SOTA. | |