Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,40 @@
+---
+license: gemma
+base_model: principled-intelligence/gemma-4-E4B-it-text-only
+language:
+  - en
+  - vi
+tags:
+  - gguf
+  - llama.cpp
+  - gemma4
+  - vocab-pruned
+pipeline_tag: text-generation
+---
+
+# Gemma-4 E4B (EN+VI vocab-pruned) — Q8_0 GGUF
+
+Vocabulary-pruned build of [principled-intelligence/gemma-4-E4B-it-text-only](https://huggingface.co/principled-intelligence/gemma-4-E4B-it-text-only): the token vocabulary is reduced to 69,246 entries covering English + Vietnamese, shrinking the embedding/LM-head so the q8_0 model fits comfortably on a 12 GB consumer GPU.
+
+- **Format**: GGUF, `q8_0` (5.2 GB)
+- **Architecture**: Gemma-4 E4B text-only — 42 layers, hidden 2560, per-layer input embeddings, sliding+full hybrid attention
+- **Vocab**: 69,246 (EN+VI pruned; original tokenizer entries outside the keep-set removed)
+- **Context**: 131,072 tokens
+- **Chat template**: Gemma-4 `<|turn>` format — use the bundled `gemma4_chat.jinja` (do not use the default Gemma 3 template)
+
+## Serving with llama.cpp
+
+```bash
+llama-server -m gemma4-e4b-envi-pruned-q8_0.gguf \
+  --host 0.0.0.0 --port 8080 -ngl 99 -c 131072 -np 16 \
+  -fa on --cache-reuse 256 --jinja \
+  --chat-template-file gemma4_chat.jinja
+```
+
+On an RTX 3060 12 GB this serves at roughly 60 tok/s single-stream.
+
+## Caveats
+
+- Text-only (vision tower removed in the upstream text-only base).
+- Because the vocab is pruned, tokenization of languages other than English/Vietnamese degrades (unknown tokens fall back to byte pieces).
+- Requires the bundled chat template; Gemma-3-style `start_of_turn` templates will not work.