exp script: bench-v2 exclusions, SAK_SEED/SAK_EPOCHS/SAK_TAG controls

job-0.5b-exp.py

@@ -32,13 +32,16 @@ from trl import SFTTrainer, SFTConfig
 
 MODE = os.environ.get("SAK_MODE", "lora-masked")
 assert MODE in ("lora-masked", "full-masked"), MODE
-random.seed(20260729)
+SEED = int(os.environ.get("SAK_SEED", "20260729"))
+EPOCHS = int(os.environ.get("SAK_EPOCHS", "3"))
+TAG = os.environ.get("SAK_TAG", "v2")
+random.seed(SEED)
 
 USER, BASE_MODEL = "Nanthasit", "Qwen/Qwen2.5-0.5B-Instruct"
 OLD_MERGED = f"{USER}/sakthai-context-0.5b-merged"
-OUT_REPO   = f"{USER}/sakthai-context-0.5b-exp-{MODE}"
+OUT_REPO   = f"{USER}/sakthai-context-0.5b-exp-{MODE}-{TAG}"
 DATASET    = f"{USER}/sakthai-combined-v7"
-BENCH      = f"{USER}/sakthai-bench-v1"
+BENCH      = f"{USER}/sakthai-bench-v2"
 EXCLUDE_URL = f"https://huggingface.co/datasets/{BENCH}/resolve/main/train_exclude_fingerprints.json"
 MAX_LEN, MAX_TURNS_PER_CONV, PARALLEL_OVERSAMPLE = 1536, 4, 3
 
@@ -140,13 +143,13 @@ else:
     lr = 2e-5
     print("full fine-tune: all", sum(p.numel() for p in model.parameters()), "params trainable")
 
-args = SFTConfig(output_dir=f"out-{MODE}", num_train_epochs=3,
+args = SFTConfig(output_dir=f"out-{MODE}-{TAG}", num_train_epochs=EPOCHS, seed=SEED,
                  per_device_train_batch_size=32, gradient_accumulation_steps=2,
                  learning_rate=lr, gradient_checkpointing=False,
                  lr_scheduler_type="cosine", warmup_ratio=0.1, logging_steps=5,
                  save_strategy="no", bf16=True, max_length=MAX_LEN,
                  completion_only_loss=True,          # mask the prompt
-                 push_to_hub=False, report_to="none", run_name=f"sakthai-0.5b-{MODE}")
+                 push_to_hub=False, report_to="none", run_name=f"sakthai-0.5b-{MODE}-{TAG}")
 trainer = SFTTrainer(model=model, args=args, train_dataset=train_ds, processing_class=tokenizer)
 trainer.train()