julienp79/occitan-gemma-3-4b-it-rslora-sfttrainer:Q5_K_M

julienp79/ occitan-gemma-3-4b-it-rslora-sfttrainer:Q5_K_M

2 Downloads Updated yesterday

Gemma 3 4B Instruct fine-tuned on Occitan (Lengadocian) via RS-LoRA (r=32). Good balance of quality and speed. Q2_K to f16 quants.

ollama run julienp79/occitan-gemma-3-4b-it-rslora-sfttrainer:Q5_K_M

curl http://localhost:11434/api/chat \
  -d '{
    "model": "julienp79/occitan-gemma-3-4b-it-rslora-sfttrainer:Q5_K_M",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='julienp79/occitan-gemma-3-4b-it-rslora-sfttrainer:Q5_K_M',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'julienp79/occitan-gemma-3-4b-it-rslora-sfttrainer:Q5_K_M',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Details

Updated yesterday

yesterday

945b3fa26e35 · 2.8GB ·

model

archgemma3

·

parameters3.88B

·

quantizationQ5_K_M

2.8GB

system

Ès un escritor e grammarista occitan lengadocian. Respon unicament en occitan lengadocian. Escriu d

265B

params

{ "num_ctx": 768, "stop": [ "<end_of_turn>" ], "temperature": 0.7, "top_

93B

template

{{- $systemPromptAdded := false }} {{- range $i, $_ := .Messages }} {{- $last := eq (len (slice $.Me

476B

Readme

Occitan Lengadocian — Gemma 3 4B RS-LoRA

Fine-tune of Gemma 3 4B Instruct on Occitan in the Lengadocian dialect (IEO grafia classica norm). Trained with pure HuggingFace PEFT — no Unsloth — which produces cleaner gradient flow at this block size.

Best 4B model in the series. Good balance of journalistic quality and literary register in a lightweight package.

Key results

Journalistic summary: 7.⁷⁄₁₀
Literary dialect consistency: 4.³⁄₅
Verb paradigm metalanguage: 4.7

System prompt

Ès un escritor e grammarista occitan lengadocian. Respon unicament en
occitan lengadocian. Escriu dirèctament lo tèxte demandat, sens cap
d'introduccion, de comentari ni d'explicacion sus ton trabalh. Pas de
preamble. Pas de version multiplas. Pas de traduccion.

Recommended quantisation

Quant	Size	Use case
Q4_K_M	~2.5 GB	Recommended — runs comfortably on CPU
Q5_K_M	~3.0 GB	Slightly better quality
Q8_0	~4.5 GB	Near-lossless
Q2_K	~1.6 GB	Minimal RAM setups

Training

RS-LoRA · r=32 · α=32 · block_size=384 · 2535 steps · 5 epochs
RTX 3060 12GB · ~4h · AutoModelForCausalLM + SFTTrainer (pure HF PEFT)

## Occitan Lengadocian — Gemma 3 4B RS-LoRA

Fine-tune of Gemma 3 4B Instruct on Occitan in the Lengadocian dialect
(IEO grafia classica norm). Trained with pure HuggingFace PEFT — no
Unsloth — which produces cleaner gradient flow at this block size.

**Best 4B model in the series.** Good balance of journalistic quality
and literary register in a lightweight package.

### Key results
- Journalistic summary: 7.7/10
- Literary dialect consistency: 4.3/5
- Verb paradigm metalanguage: 4.7

### System prompt
```
Ès un escritor e grammarista occitan lengadocian. Respon unicament en
occitan lengadocian. Escriu dirèctament lo tèxte demandat, sens cap
d'introduccion, de comentari ni d'explicacion sus ton trabalh. Pas de
preamble. Pas de version multiplas. Pas de traduccion.
```

### Recommended quantisation
| Quant | Size | Use case |
|-------|------|----------|
| Q4_K_M | ~2.5 GB | Recommended — runs comfortably on CPU |
| Q5_K_M | ~3.0 GB | Slightly better quality |
| Q8_0 | ~4.5 GB | Near-lossless |
| Q2_K | ~1.6 GB | Minimal RAM setups |

### Training
RS-LoRA · r=32 · α=32 · block_size=384 · 2535 steps · 5 epochs  
RTX 3060 12GB · ~4h · AutoModelForCausalLM + SFTTrainer (pure HF PEFT)

Paste, drop or click to upload images (.png, .jpeg, .jpg, .svg, .gif)