AtomicChat/embeddinggemma-2-GGUF-metrics

embeddinggemma-2 GGUF, everything behind the numbers

This dataset holds the measurements, logs and inputs behind AtomicChat/embeddinggemma-2-GGUF.

What is here

Path	What it is
results.json	Every measured file: ours, Unsloth's, AutoRound's (webmp3/Sakura-EmbeddingGemma-2-AutoRound-GGUF) and ggml-org's Q8_0. Each row has the size and, per eval set and width (768, 256), the mean and p99 of 1 - cosine to BF16, the same-top-result rate with its 95% interval, top-10 overlap and recall@1
imatrix/imatrix.gguf	The importance matrix (2,425 chunks of 512 tokens, 216 tensors)
imatrix/calib.txt, calib.txt.manifest.json	The calibration text and where each part came from (source file hashes, document ids)
evalset/set.json	The 1,400 held-out query and passage pairs (calib-corpora eval/neutral and eval/code)
reference/embeddinggemma-2-BF16.npz	BF16 vectors for every query and passage, to measure your own files against
experiments/*.json	The layout search: one tensor group at a time, type curves, depth bands, candidates. experiments/paired.json holds the head-to-head comparisons with Unsloth and AutoRound on the same texts, from scripts/paired_compare.py
mm/*.json	The projector sanity check (one image, one audio clip)
logs/	Conversion, --check-tensors, the tensor-for-tensor comparisons with ggml-org and Unsloth, imatrix and its statistics, every quantize run, the real tensor types of each file, the llama.cpp commit
scripts/	The calibration builder, the evaluation harness and the GGUF checks

How the measurements were made

scripts/embed_eval.py starts llama-server --embeddings --pooling mean -fa off for each file and embeds every query (task: search result | query: ..., or task: code retrieval | query: ...) and passage (title: ... | text: ...). It then compares the vectors with the BF16 reference:

• the cosine for the same text;
• whether each query's best passage, out of every passage in the set, is the same as with the reference;
• intervals by bootstrap over documents, because passages from one document are not independent.

We ran it on an Apple M4 Max with llama.cpp 26908739bc8a. The BF16 file run twice gives identical vectors.

Why the layouts look like this

experiments/ has the numbers. Measured one tensor group at a time against Q8_0:

• the embedding table token_embd is the cheapest place to save bytes, about six times cheaper per megabyte than the transformer layers;
• the output head is the most fragile tensor;
• blocks 0-5 and 18-23 are 2-3 times more sensitive than the middle.

The ratios also added up: the per-band costs sum to the cost of the whole body. So each file's layout was picked by adding up the measured costs within Unsloth's size, and the best candidates were then measured whole.