Merge pull request #18 from atlasia-ma/fix/review-feedback-training

address review feedback

Author: hBouanane

pyproject.toml

@@ -11,14 +11,14 @@ dev = [
     "pytest>=9.1.1",
 ]
 eval = [
-    "sacrebleu>=2.4",
+    "sacrebleu>=2.6.0",
 ]
 train = [
-    "datasets>=4.3.0",
-    "trl>=0.24.0",
-    "unsloth>=2026.6.9",
-    "wandb>=0.19",
-    "python-dotenv>=1.0",
+    "datasets>=5.0.0",
+    "trl>=1.7.1",
+    "unsloth>=2026.7.2",
+    "wandb>=0.28.0",
+    "python-dotenv>=1.2.2",
 ]
 
 [project.scripts]

src/darija_translator/cli.py

@@ -59,11 +59,13 @@ def main():
     subparsers = parser.add_subparsers(required=True)
 
     train_parser = subparsers.add_parser("train")
-    train_parser.add_argument("--dataset", default="atlasia/darija_english")
+    train_parser.add_argument("--dataset",
+                              default="atlasia/darija-english-combined")
     train_parser.set_defaults(func=run_train)
 
     eval_parser = subparsers.add_parser("evaluate")
-    eval_parser.add_argument("--dataset", default="atlasia/darija_english")
+    eval_parser.add_argument("--dataset",
+                             default="atlasia/darija-english-combined")
     eval_parser.set_defaults(func=run_evaluate)
 
     args = parser.parse_args()

src/darija_translator/evaluate.py

@@ -32,7 +32,7 @@ def generate_translations(model, tokenizer, sources: list[str],
         ).to(model.device)
         input_len = inputs["input_ids"].shape[1]
         outputs = model.generate(**inputs,
-                                 max_new_tokens=128,
+                                 max_new_tokens=256,
                                  do_sample=False,
                                  use_cache=True)
         new_tokens = outputs[:, input_len:]

src/darija_translator/train.py

@@ -8,7 +8,8 @@
 
 def build_trainer(model, tokenizer, train_dataset, eval_dataset,
                   config: TrainConfig):
-    if config.report_to == "wandb" or (isinstance(config.report_to, list) and "wandb" in config.report_to):
+    if config.report_to == "wandb" or (isinstance(config.report_to, list)
+                                       and "wandb" in config.report_to):
         os.environ["WANDB_PROJECT"] = config.wandb_project
 
     trainer = SFTTrainer(
@@ -33,7 +34,7 @@ def build_trainer(model, tokenizer, train_dataset, eval_dataset,
             lr_scheduler_type=config.lr_scheduler_type,
             seed=config.seed,
             report_to=config.report_to,
-            group_by_length=config.group_by_length,
+            # group_by_length=config.group_by_length,
         ),
     )
     return train_on_responses_only(