TiGa-RCE's picture
Finalize reproducibility bundle and locked revisions
f3a14ee verified
|
Raw
History Blame Contribute Delete
1.96 kB

Frozen result summary

The representation gate required minimum aligned cosine of at least 0.99 against the matching BF16 reference, with no Top-1 or Recall@5 regression, MRR loss no worse than 0.01, and no more than two changed query ranks.

All 30 full-dimensional MLX checkpoints retained 24/24 Top-1 retrieval. The best result among Q, oQ, and oQe at each nominal bit width was:

Family Best Q4 Best Q6 Best Q8
Qwen3 0.6B oQ4e 0.943953 — fail oQ6e 0.987972 — fail oQ8e 0.998660 — pass
GTE-Qwen2 1.5B oQ4e 0.977315 — fail oQ6e 0.998114 — pass oQ8e 0.999547 — pass
Qwen3 8B oQ4e 0.983151 — fail oQ6e 0.998220 — pass oQ8e 0.999568 — pass

All 54 mixed BF16/quantized query-index directions retained 24/24 Top-1. Applying a threshold calibrated once on BF16 without retuning changed decisions; the largest break was five false negatives out of 24 for Qwen3 0.6B standard-Q4 queries against the BF16 index.

The CUDA-native bitsandbytes controls are not MLX format equivalents:

Family CUDA format Minimum cosine vs same-path CUDA BF16 Top-1 Gate
Qwen3 0.6B INT8 0.99257 1.0 pass
Qwen3 0.6B NF4 0.90902 1.0 fail
GTE-Qwen2 1.5B INT8 0.98239 1.0 fail
GTE-Qwen2 1.5B NF4 0.89340 1.0 fail
Qwen3 8B INT8 0.99114 1.0 pass
Qwen3 8B NF4 0.97268 1.0 fail

GTE-Qwen2 1.5B is excluded from cross-runtime representation trends. Its root-packed and direct CUDA BF16 vectors diverged, and direct CUDA BF16 also diverged from local MLX BF16. Its quantized CUDA comparisons are interpreted only against the direct same-loading-path CUDA BF16 reference.

These are deterministic measurements on one frozen 24-pair smoke set. They do not establish universal embedding quality or replace broader retrieval, clustering, similarity, threshold, and batching evaluations.