---
license: gemma
base_model: google/gemma-3-4b-it
library_name: gguf
tags:
  - gguf
  - gemma
  - gemma-3
  - cnc
  - manufacturing
  - work-order
  - ll-machinery
  - fine-tuned
  - ollama
pipeline_tag: text-generation
---

# gemma4-cnc-ll (GGUF) — Two-stage Fine-tune

Two-stage QLoRA fine-tune of [`google/gemma-3-4b-it`](https://huggingface.co/google/gemma-3-4b-it):

**Stage 1 — Work-order knowledge**
CNC work-order database schema, BOM categories, pricing models, Neo4j/SQL queries
for 永詮機械 (Yong Chuan Machinery).

**Stage 2 — L&L 永詮 specialization** (anti-dilution training)
14 machine models (LLA/LLB/LLS/LL/LFM/LFS/LS-C/LS-S/LS-U/LD/LC/TA/MA/LA/A),
6 application industries, company background. Trained with 5 anti-dilution
mechanisms including critical-facts validation.

## Quick start (Ollama)

```bash
ollama run hf.co/Pauldyu57/gemma4-cnc-ll-GGUF:Q4_K_M
```

Other tags: `Q5_K_M`, `Q8_0`, `F16`.

If this repo is private, link your Ollama SSH key first:
```bash
cat ~/.ollama/id_ed25519.pub
# → paste into https://huggingface.co/settings/keys
```

## Files

| Quant   | Approx size | Notes                        |
|---------|-------------|------------------------------|
| F16     | ~7.5 GB     | Reference                    |
| Q8_0    | ~4.3 GB     | Near-lossless                |
| Q5_K_M  | ~3.0 GB     | Good quality / size balance  |
| Q4_K_M  | ~2.5 GB     | Default for CPU / small GPUs |

## Base model

[`google/gemma-3-4b-it`](https://huggingface.co/google/gemma-3-4b-it) — usage subject to the
[Gemma Terms of Use](https://ai.google.dev/gemma/terms).