ChamalyAI/gemma4-E2B-chamaliya

Gemma4-E2B Darija North — Merged (RAFT Ready)

This is a fully merged (LoRA weights fused into base model) version of google/gemma-4-E2B-it
fine-tuned on ~60K Northern Moroccan Darija (Jebli dialect) samples.

Training

• Base model: google/gemma-4-E2B-it
• Dataset: Northern Darija dialect data (~60K samples)
• Method: LoRA fine-tuning via Unsloth on Kaggle T4x2
• LoRA rank: 16, alpha: 32
• This model: LoRA adapter merged into base → full model, no adapter needed

RAFT Ready

Because the adapter is merged, this model can be used directly as a policy model
for RAFT (Reward rAnked Fine-Tuning) without any special adapter handling.

Dialect Features

• Northern Darija / Jebli phonology: preserved Qaf (ق), not shifted to Gaf
• Pronouns: uses شمالي gender-neutral forms (نتينا)
• Spanish loanwords: النيبيرا (nevera), etc.
• Diminutives: عايل ستيتو / عايلة ستيتوة
• Time expressions: القايلة, etc.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("ChamalyAI/gemma4-E2B-chamaliya", torch_dtype="auto")
tokenizer = AutoTokenizer.from_pretrained("ChamalyAI/gemma4-E2B-chamaliya")