Views
No views yet
1import torch
2from peft import PeftModel
3from transformers import AutoModelForCausalLM, AutoTokenizer
4
5# Load base model
6base_model = AutoModelForCausalLM.from_pretrained(
7 "CohereLabs/aya-expanse-8b",
8 device_map="auto",
9 trust_remote_code=True
10)
11
12# Load Tunisian adapter
13model = PeftModel.from_pretrained(
14 base_model,
15 "alabenayed/improved-aya-expanse-8b-cpt-tunisian",
16 device_map="auto"
17)
18model.eval()
19
20# Load tokenizer
21tokenizer = AutoTokenizer.from_pretrained(
22 "alabenayed/improved-aya-expanse-8b-cpt-tunisian",
23 trust_remote_code=True
24)
25if tokenizer.pad_token is None:
26 tokenizer.pad_token = tokenizer.eos_token
27
28# Generate with chat template
29def generate_prompt(prompt: str, max_new_tokens: int = 160):
30 messages = [{"role": "user", "content": prompt}]
31 inputs = tokenizer.apply_chat_template(
32 messages,
33 add_generation_prompt=True,
34 tokenize=True,
35 return_dict=True,
36 return_tensors="pt",
37 )
38
39 device = next(model.parameters()).device
40 inputs = {k: v.to(device) for k, v in inputs.items()}
41
42 outputs = model.generate(
43 **inputs,
44 max_new_tokens=max_new_tokens,
45 temperature=0.7,
46 do_sample=True,
47 )
48
49 return tokenizer.decode(outputs[0], skip_special_tokens=True)
50
51# Try with Tunisian prompts
52prompts = [
53 "عسلامة تحكي تونسي ؟", # How are you speaking Tunisian?
54 "اليوم الطقس مزيان شنوة نجم نعمل؟", # Today's weather is nice, what can we do?
55 "اش نجم نقول لصاحبي باش نهنيه بعيد ميلادو؟", # What can I say to my friend on his birthday?
56]
57
58for prompt in prompts:
59 output = generate_prompt(prompt)
60 print(f"📝 {prompt}\n💬 {output}\n")r=16 (rank)α=32 (scaling)dropout=0.05| Step | Loss | Eval Loss |
|---|---|---|
| 0 | 2.500 | - |
| 1,000 | 2.260 | 2.214 |
| 2,000 | 2.194 | 2.193 |
| 3,000 | 2.150 | 2.132 |
| 4,000 | 2.123 | 2.098 |
| 5,113 | 2.105 | 2.089 |
1@misc{tunisian-llm-2026,
2 author = {MohamedAla-Benayed},
3 title = {Improved Aya Expanse 8B CPT Tunisian},
4 year = {2026},
5 publisher = {Hugging Face},
6 howpublished = {\url{https://huggingface.co/alabenayed/improved-aya-expanse-8b-cpt-tunisian}}
7}