Update README.md

README.md

@@ -1,3 +1,23 @@
----
-license: gemma
----
+---
+license: gemma
+metrics:
+- perplexity
+base_model:
+- google/gemma-3-1b-it-qat-q4_0-gguf
+- bartowski/google_gemma-3-1b-it-GGUF
+---
+
+This is a "self" merge of https://huggingface.co/google/gemma-3-1b-it-qat-q4_0-gguf and https://huggingface.co/bartowski/google_gemma-3-1b-it-GGUF.
+
+The official QAT weights released by google use fp16 (instead of Q6_K) for the embeddings table, which makes this model take a significant extra amount of memory (and storage) compared to what Q4_0 quants are supposed to take. Instead of quantizing the table myself, I extracted it from Bartowski's quantized models, because those were already calibrated with imatrix, which should squeeze some extra performance out of it.
+
+Here are some perplexity measurements:
+
+| Model | File size ↓ | PPL (wiki.text.raw) ↓ |
+| --- | --- | --- |
+| [This model](https://huggingface.co/stduhpf/google-gemma-3-1b-it-qat-q4_0-gguf-small/blob/main/gemma-3-1b-it-q4_0_s.gguf) | 720 MB | 28.2603 +/- 0.26947 |
+| [Q4_0 (bartowski)](https://huggingface.co/bartowski/google_gemma-3-1b-it-GGUF/blob/main/google_gemma-3-1b-it-Q4_0.gguf) | 722 MB | 34.4906 +/- 0.34539 |
+| [QAT Q4_0 (google)](https://huggingface.co/google/gemma-3-1b-it-qat-q4_0-gguf/blob/main/gemma-3-1b-it-q4_0.gguf) | 1 GB | 28.2603 +/- 0.26947 |
+
+Note that this model ends up smaller than the Q4_0 from Bartowski. This is because llama.cpp sets some tensors to Q4_1 when quantizing models to Q4_0, but Google decided to use only Q4_0 instead, which is slightly smaller.
+The perplexity scores are barely within margin of error between this model and the original QAT, it seems like the embedding table starts making a difference at this small size, though the trade off is probably still worth it..