Add model card

README.md

@@ -0,0 +1,103 @@
+---
+license: gemma
+base_model: lapa-llm/lapa-v0.1.3-instruct
+language:
+- uk
+tags:
+- gguf
+- llama.cpp
+- ukrainian
+- gemma3
+- quantized
+- q8_0
+library_name: gguf
+pipeline_tag: text-generation
+---
+
+# Lapa v0.1.3 Instruct — Q8_0 GGUF
+
+8-bit (`Q8_0`) [GGUF](https://github.com/ggml-org/llama.cpp) quantization of
+[**lapa-llm/lapa-v0.1.3-instruct**](https://huggingface.co/lapa-llm/lapa-v0.1.3-instruct) —
+a Ukrainian-specialized instruction model built on Google's Gemma 3 12B, with roughly
+80 000 of its 256 000 tokenizer tokens replaced by Ukrainian-oriented ones.
+
+This repository contains **only the quantized weights**. For the original model, tokenizer
+sources, training details and intended use, see the
+[upstream model card](https://huggingface.co/lapa-llm/lapa-v0.1.3-instruct).
+
+## Files
+
+| File | Quant | Size | Notes |
+|------|-------|------|-------|
+| `lapa-v0.1.3-instruct-Q8_0.gguf` | `Q8_0` | ~12.5 GB | Near-lossless 8-bit; fits a single 24 GB GPU |
+
+## Why Q8_0
+
+`Q8_0` is the highest-fidelity common GGUF quant (8-bit, ~0.5 GB per B params). It keeps
+generation quality effectively indistinguishable from the `bf16` original while roughly
+halving memory, so the full 12B model fits comfortably on one 24 GB card (e.g. RTX 3090/4090)
+with room for KV cache. This matters here because Lapa's **modified tokenizer** makes the
+model sensitive to precision loss — lower-bit quants (Q4/Q5/Q6) risk degrading Ukrainian
+morphology, so Q8_0 is the recommended operating point.
+
+## How it was made
+
+Following the [official Lapa quantization recipe](https://github.com/lapa-llm/lapa-llm/tree/main/quantize),
+which uses a patched converter that preserves Lapa's custom tokenizer (`tokenizer.model` +
+`merge_info.json`) so no vocabulary precision is lost:
+
+```bash
+# llama.cpp @ b8185
+# 1. HF safetensors -> F16 GGUF (patched Lapa converter)
+python convert_hf_to_gguf_lapa.py lapa-v0.1.3-instruct/ \
+    --outfile lapa-12B-it-F16.gguf --outtype f16
+
+# 2. F16 -> Q8_0
+llama-quantize lapa-12B-it-F16.gguf lapa-v0.1.3-instruct-Q8_0.gguf Q8_0
+```
+
+## Usage
+
+### `llama-server` (OpenAI-compatible API)
+
+```bash
+llama-server \
+  --model lapa-v0.1.3-instruct-Q8_0.gguf \
+  --host 0.0.0.0 --port 8080 \
+  --n-gpu-layers 999 --ctx-size 8192 --flash-attn on
+```
+
+```bash
+curl http://127.0.0.1:8080/v1/chat/completions \
+  -H 'Content-Type: application/json' \
+  -d '{"messages":[{"role":"user","content":"Розкажи коротко про місто Львів."}],
+       "temperature":0.3}'
+```
+
+### `llama-cli`
+
+```bash
+llama-cli -m lapa-v0.1.3-instruct-Q8_0.gguf -ngl 999 -c 8192 \
+  -p "Перепиши речення без цифр: На 500 сторінках є примітки."
+```
+
+### Download
+
+```bash
+huggingface-cli download skypro1111/lapa-v0.1.3-instruct-Q8_0-GGUF \
+  lapa-v0.1.3-instruct-Q8_0.gguf --local-dir .
+```
+
+## License
+
+Derived from Google Gemma 3 and inherits the [**Gemma Terms of Use**](https://ai.google.dev/gemma/terms).
+Use of these weights is subject to those terms and to any terms of the upstream
+[lapa-llm/lapa-v0.1.3-instruct](https://huggingface.co/lapa-llm/lapa-v0.1.3-instruct) model.
+
+## Credits
+
+- Original model: [**Lapa LLM**](https://huggingface.co/lapa-llm) team — `lapa-llm/lapa-v0.1.3-instruct`
+- Base architecture: [Google Gemma 3](https://ai.google.dev/gemma)
+- Quantization tooling: [llama.cpp](https://github.com/ggml-org/llama.cpp)
+
+*This is an unofficial, community quantization. Not affiliated with the Lapa LLM team or Google.*