adaption_moroccan_darija_prompts_tril

adaption_moroccan_darija_prompts_tril

Model Training

A LORA adapter for Qwen/Qwen3.5-0.8B. This model was trained with SFT using Adaption's AutoScientist on the moroccan_darija_prompts & trilingual_codeswitch_chat dataset.

AutoScientist Config

{
  "job_id": "7f82f704-836d-4ce6-884b-a110165e7759",
  "training_experiment_id": "904b5d37-2f32-4f4f-a70b-d609f837ff7e",
  "original_model_name": "Qwen/Qwen3.5-0.8B",
  "trained_model_name": "adaption_moroccan_darija_prompts_tril",
  "training_method": "sft",
  "training_type": "lora",
  "data_format": "chat",
  "hyperparams": {
    "lora": "true",
    "lora_r": 32,
    "n_evals": 5,
    "n_epochs": 5,
    "batch_size": "max",
    "lora_alpha": 128,
    "lora_dropout": 0,
    "min_lr_ratio": 0.1,
    "warmup_ratio": 0.08,
    "weight_decay": 0.01,
    "learning_rate": 0.0001,
    "max_grad_norm": 1,
    "base_model_size": "0.8B",
    "train_on_inputs": "false",
    "training_method": "sft",
    "lr_scheduler_type": "cosine",
    "scheduler_num_cycles": 0.5,
    "lora_trainable_modules": "q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj"
  }
}

Training Data

The model was trained on 7,429 rows of adapted data with the following domain distribution: other (17%), language (8%), travel (8%), medical (5%), religion (5%), transportation (4%), cooking (4%), product-advice (4%), culture (3%), academic-education (3%), geography (3%), how-to (3%), personal-finance (2%), science (2%), personal-growth (2%), history (2%), technology (2%), entertainment (2%), sports (1%), dating (1%), math (1%), fashion-beauty (1%), animal-nature (1%), parenting-family (1%), corporate-business (1%), legal (1%), music (1%), governance (1%), real-estate (1%), code (1%), news (1%), fitness-sports (1%), art (1%), career-workplace (1%), agriculture (1%), customer-service (1%), writing-editing-communication (1%), beauty (1%), games (0%), architecture-design (0%), writing-edit游戏副本-communication (0%), literature (0%), social (0%), roleplay (0%), philosophy (0%), education (0%), market-analysis (0%), finance (0%), customer-support (0%), marketing (0%), data-analysis-visualization (0%), hr (0%), logic (0%), economics (0%), psychology (0%), media (0%), energy (0%), politics (0%), philosophy-ethics (0%), survival-outdoors (0%), govern游戏副本 (0%), academic-游戏副本 (0%), business (0%), lifestyle (0%), mental-health (0%), home-lifestyle (0%), urban-planning (0%), logic-reasoning (0%), community-events (0%), career-work-place (0%), outdoor-recreation (0%), community-engagement (0%), community-development (0%), survival-preparedness (0%), food (0%), coffee (0%), direct (0%), ethics (0%), health (0%), safety (0%), open_qa (0%), mythology (0%), nutrition (0%), accounting (0%), it-support (0%), govern-ance (0%), integration (0%), photography (0%), govern ance (0%), communication (0%), food-beverage (0%), event-planning (0%), social-justice (0%), personal-ethics (0%), product-support (0%), survival-skills (0%), technical-issue (0%), humanitarian-aid (0%), philosophy-logic (0%), agric游戏副本 (0%), business-strategy (0%), community-support (0%), currency-exchange (0%), logical-reasoning (0%), urban-development (0%), academia-education (0%), ethical-philosophy (0%), recreation-leisure (0%), academics-education (0%), community-management (0%), community-organizing (0%), family-relationships (0%), lifestyle-activities (0%), community-involvement (0%), reading-comprehension (0%), community-organization (0%), personal-privacy-ethics (0%), writing-edit grinding-communication (0%), writing-edit conflated with roleplay (0%).

Model Evaluation

The model was evaluated on an in-distribution held-out test set as well as a broader domain-specific test set to measure generalization.

How to use

pip install torch transformers peft

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE = "Qwen/Qwen3.5-0.8B"
ADAPTER = "<this-repo-id>"

device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float32 if device == "cpu" else torch.bfloat16

base = AutoModelForCausalLM.from_pretrained(BASE, dtype=dtype).to(device)
model = PeftModel.from_pretrained(base, ADAPTER)
# Optional: merge the LoRA weights into the base for faster inference
model = model.merge_and_unload()
model.eval()

tokenizer = AutoTokenizer.from_pretrained(BASE)
messages = [{"role": "user", "content": "Hello!"}]
text = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(device)

with torch.inference_mode():
    out = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))