Upload sft_03_train.py with huggingface_hub

sft_03_train.py

@@ -179,8 +179,10 @@ def sample_test(model, tokenizer, prompt_text: str, max_new: int = 200):
     with torch.no_grad():
         with torch.amp.autocast(device_type="cuda", dtype=DTYPE):
             out = model.generate(x, max_new_tokens=max_new,
-                                  temperature=0.7, top_k=50,
-                                  repetition_penalty=1.1)
+                                  temperature=0.3, top_k=40,
+                                  repetition_penalty=1.0,
+                                  no_repeat_ngram_size=0,
+                                  eos_token_id=0)  # <|endoftext|>
     text = tokenizer.decode(out[0].tolist())
     model.train()
     return text
@@ -406,9 +408,14 @@ def main():
             try:
                 mdl = model._orig_mod if hasattr(model, "_orig_mod") else model
                 prompts = [
-                    "### Kullanici:\n2 + 2 kac eder?\n### Asistan:\n",
+                    # Tarif (training'de turkish_recipes 4K)
+                    "### Kullanici:\nMercimek corbasi tarifi ver.\n### Asistan:\n",
+                    # Factual short (turkish_exam, knowledge)
                     "### Kullanici:\nTurkiye'nin baskenti neresidir?\n### Asistan:\n",
-                    "### Kullanici:\nBir cocuga uzay neden onemli oldugunu anlat.\n### Asistan:\n",
+                    # Aciklayici (general)
+                    "### Kullanici:\nBir e-mail nasil yazilir? Kisa anlat.\n### Asistan:\n",
+                    # Simple math (gsm8k/metamath)
+                    "### Kullanici:\nAhmet'in 5 elmasi var, 2 tane yer. Kac kaldi?\n### Asistan:\n",
                 ]
                 for p in prompts:
                     out = sample_test(mdl, args.tokenizer, p, max_new=120)