Add Q5_K_S after text embedding smoke pass

.gitattributes

@@ -43,3 +43,4 @@ EmbeddingGemma-2-Q2_K.gguf filter=lfs diff=lfs merge=lfs -text
 EmbeddingGemma-2-IQ2_XS.gguf filter=lfs diff=lfs merge=lfs -text
 EmbeddingGemma-2-IQ1_M.gguf filter=lfs diff=lfs merge=lfs -text
 EmbeddingGemma-2-Q1_0.gguf filter=lfs diff=lfs merge=lfs -text
+EmbeddingGemma-2-Q5_K_S.gguf filter=lfs diff=lfs merge=lfs -text

EmbeddingGemma-2-Q5_K_S.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c0dc8e88bc69c720f997ea1dd9cf4a57bbbccaecfa85925455a6608654d2d81
+size 210670208

README.md

@@ -31,7 +31,7 @@ Thank you for supporting this work.
 
 ## Embedding evaluation
 
-All values below were measured locally from the locked BF16 GGUF and quantized artifacts; they are not copied from the upstream model card. The fixed hold-out evaluation uses the 1,379-pair test split of [MTEB STSBenchmark STS](https://huggingface.co/datasets/mteb/stsbenchmark-sts), dataset revision `96943a16ea6a35129e253c659081cb59daf81b30`. It covers 2,552 unique sentences with the `task: sentence similarity | query:` prefix, an 8,192-token context, and the CPU `llama.cpp` runtime at commit `9c2e0e491a822adae1f0b1c831adb4160057d24f`. Spearman and Pearson measure correlation between cosine similarity and the human scores. Mean and 5th-percentile cosine measure vector agreement with the BF16 GGUF reference. Higher values indicate stronger task correlation or closer vector agreement; smaller files use less disk. This is a task-specific embedding evaluation, not the full MTEB benchmark. Token-level KLD, next-token Top-1, PPL, and RMS probability deltas do not apply to this embedding model.
+All values below were measured locally from the locked BF16 GGUF and quantized artifacts; they are not copied from the upstream model card. Additional quant formats are being added incrementally and will appear in the comparison table after running this same hold-out evaluation. The fixed hold-out evaluation uses the 1,379-pair test split of [MTEB STSBenchmark STS](https://huggingface.co/datasets/mteb/stsbenchmark-sts), dataset revision `96943a16ea6a35129e253c659081cb59daf81b30`. It covers 2,552 unique sentences with the `task: sentence similarity | query:` prefix, an 8,192-token context, and the CPU `llama.cpp` runtime at commit `9c2e0e491a822adae1f0b1c831adb4160057d24f`. Spearman and Pearson measure correlation between cosine similarity and the human scores. Mean and 5th-percentile cosine measure vector agreement with the BF16 GGUF reference. Higher values indicate stronger task correlation or closer vector agreement; smaller files use less disk. This is a task-specific embedding evaluation, not the full MTEB benchmark. Token-level KLD, next-token Top-1, PPL, and RMS probability deltas do not apply to this embedding model.
 
 ### BF16 reference baseline
 

SHA256SUMS.txt

@@ -5,6 +5,7 @@ c4e59a215a94e4ab077e1a10c43df85ab1c63addc8fdf5515a5ee9b99c8a15f4  EmbeddingGemma
 a5c902b15458c4886d6f1296bad71fd8084d7ac626cb8713778cf2af5e1b5a71  EmbeddingGemma-2-Q3_K_M.gguf
 ea25550b050fa8826707c5e44659a7c18259606dcc6ba449f771e3c75cf4e663  EmbeddingGemma-2-Q4_K_M.gguf
 22f79226a02bfa2dd840414f342a004f93b3cb85e498923a5e089f03f6410b1f  EmbeddingGemma-2-Q5_K_M.gguf
+0c0dc8e88bc69c720f997ea1dd9cf4a57bbbccaecfa85925455a6608654d2d81  EmbeddingGemma-2-Q5_K_S.gguf
 21693ca41877ffb20727763e167c503b0467eac4ccef51551230d72c2afcd37e  EmbeddingGemma-2-Q6_K.gguf
 eec0ccc2e76aa85b415651e869c25e476279d6b7d1985533c8296f13989efa10  EmbeddingGemma-2-Q8_0.gguf
 b663b6822d1f2d641acc54fc0adfc143b630b7ac87ec54fd60888f325071a0de  mmproj-EmbeddingGemma-2-BF16.gguf

reproducibility/artifact-measurements.tsv

@@ -9,3 +9,4 @@ EmbeddingGemma-2-Q2_K.gguf	120394368	0.120394	PASS_EMBEDDING_CPU
 EmbeddingGemma-2-IQ2_XS.gguf	110945920	0.110946	PASS_EMBEDDING_CPU
 EmbeddingGemma-2-IQ1_M.gguf	103040640	0.103041	PASS_EMBEDDING_CPU
 EmbeddingGemma-2-Q1_0.gguf	66163328	0.066163	PASS_EMBEDDING_CPU
+EmbeddingGemma-2-Q5_K_S.gguf	210670208	0.210670	PASS_EMBEDDING_CPU

reproducibility/manifest.md

@@ -105,3 +105,10 @@ The fixed task evaluation uses the full MTEB STSBenchmark `test` split and the u
 - The measured memory/fidelity candidates are `EmbeddingGemma-2-Q5_K_M.gguf` and `EmbeddingGemma-2-Q4_K_M.gguf`: both retain STS correlations close to BF16 while reducing file size to `0.212759 GB` and `0.181695 GB`. No VRAM or throughput claim is made. `Q1_0` returns valid vectors but has Spearman `0.061457` and mean cosine agreement `0.513763`, indicating severe loss on this task.
 - Machine-readable results: [`embedding-evaluation.tsv`](embedding-evaluation.tsv). The local evaluation runner is [`evaluate-stsbenchmark.py`](evaluate-stsbenchmark.py); raw test data and runtime logs remain local.
 - Evaluation ran in a CPU-only systemd cgroup with `MemoryMax=32G` and `MemorySwapMax=0`; cgroup peak memory was `1060462592` bytes and swap peak was `0`. Local observed MemAvailable ranged from `247175248` kB at start to a sampled minimum of `242637836` kB. Resource details are kept in the local `reports/embeddinggemma-2/evaluation-resource.tsv`.
+
+## Incremental experiment: Q5_K_S
+
+- Artifact: `EmbeddingGemma-2-Q5_K_S.gguf`; SHA256 `0c0dc8e88bc69c720f997ea1dd9cf4a57bbbccaecfa85925455a6608654d2d81`; decimal size `0.210670 GB`.
+- Quantization command: `llama-quantize --imatrix calibration/embeddinggemma-2/embeddinggemma-2-sts.imatrix.gguf embeddinggemma-2-text-BF16.gguf EmbeddingGemma-2-Q5_K_S.gguf Q5_K_S 8`.
+- CPU `llama-server --embedding --pooling mean -ngl 0` smoke passed with finite normalized 768-dimensional text vectors.
+- Build cgroup: `MemoryMax=32G`, `MemorySwapMax=0`; peak memory `974057472` bytes and swap peak `0` bytes.