Upload 2 files

training_infos.md

@@ -0,0 +1,72 @@
+# Promt Format
+```
+alpaca_prompt = """Sen bir doktorsun. Soruları buna göre cevapla.
+### Soru:
+{}
+
+### Cevap:
+{}"""
+```
+
+# Training args
+```
+batch_size = 128
+gradient_accumulation_steps = 8
+num_train_epochs = 2
+per_device_batch_size = int(batch_size / gradient_accumulation_steps)
+training_args = TrainingArguments(
+        per_device_train_batch_size = per_device_batch_size,
+        per_device_eval_batch_size = per_device_batch_size,
+        gradient_accumulation_steps = gradient_accumulation_steps,
+        save_total_limit = 1,
+        warmup_steps = int(2000 / batch_size),
+        num_train_epochs = num_train_epochs,
+        learning_rate = 1e-4,
+        fp16 = not is_bfloat16_supported(),
+        bf16 = is_bfloat16_supported(),
+        optim = "adamw_8bit",
+        weight_decay = 0.01,
+        lr_scheduler_type = "linear",
+        seed = 3407,
+        output_dir = output_dir,
+        save_strategy = "steps",
+        eval_strategy = "steps",
+        logging_strategy = "steps",
+        save_steps = int(5000 / batch_size * num_train_epochs),
+        eval_steps = int(28900 / batch_size * num_train_epochs),
+        logging_steps = int(28900 / batch_size * num_train_epochs),
+    )
+```
+
+# Trainer args
+
+```
+max_seq_length = 8192
+trainer = SFTTrainer(
+    model = model,
+    tokenizer = tokenizer,
+    train_dataset = train_dataset,
+    eval_dataset = eval_dataset,
+    dataset_text_field = "text",
+    max_seq_length = max_seq_length,
+    dataset_num_proc = 1,
+    packing = False, # Can make training 5x faster for short sequences.
+    args = training_args
+)
+```
+
+# From pretrained args
+
+```
+from unsloth import FastLanguageModel
+max_seq_length = 8192
+dtype = None
+load_in_4bit = False
+
+model, tokenizer = FastLanguageModel.from_pretrained(
+    model_name = output_dir,
+    max_seq_length = max_seq_length,
+    dtype = dtype,
+    load_in_4bit = load_in_4bit,
+)
+```

training_log.json

@@ -0,0 +1,146 @@
+[
+    {
+        "loss": 2.0908,
+        "grad_norm": 0.5866689682006836,
+        "learning_rate": 8.923190911336132e-05,
+        "epoch": 0.22189421894218941,
+        "step": 451
+    },
+    {
+        "eval_loss": 1.9633848667144775,
+        "eval_runtime": 1103.6079,
+        "eval_samples_per_second": 26.193,
+        "eval_steps_per_second": 1.637,
+        "epoch": 0.22189421894218941,
+        "step": 451
+    },
+    {
+        "loss": 1.9265,
+        "grad_norm": 0.5757979154586792,
+        "learning_rate": 7.809335638429242e-05,
+        "epoch": 0.44378843788437883,
+        "step": 902
+    },
+    {
+        "eval_loss": 1.9037492275238037,
+        "eval_runtime": 1102.9066,
+        "eval_samples_per_second": 26.21,
+        "eval_steps_per_second": 1.638,
+        "epoch": 0.44378843788437883,
+        "step": 902
+    },
+    {
+        "loss": 1.8852,
+        "grad_norm": 0.5880784392356873,
+        "learning_rate": 6.695480365522352e-05,
+        "epoch": 0.6656826568265682,
+        "step": 1353
+    },
+    {
+        "eval_loss": 1.8703107833862305,
+        "eval_runtime": 1103.8364,
+        "eval_samples_per_second": 26.188,
+        "eval_steps_per_second": 1.637,
+        "epoch": 0.6656826568265682,
+        "step": 1353
+    },
+    {
+        "loss": 1.8585,
+        "grad_norm": 0.6274667978286743,
+        "learning_rate": 5.581625092615461e-05,
+        "epoch": 0.8875768757687577,
+        "step": 1804
+    },
+    {
+        "eval_loss": 1.8478941917419434,
+        "eval_runtime": 1103.7709,
+        "eval_samples_per_second": 26.189,
+        "eval_steps_per_second": 1.637,
+        "epoch": 0.8875768757687577,
+        "step": 1804
+    },
+    {
+        "loss": 1.8051,
+        "grad_norm": 0.6508978009223938,
+        "learning_rate": 4.4677698197085704e-05,
+        "epoch": 1.1094710947109472,
+        "step": 2255
+    },
+    {
+        "eval_loss": 1.835593819618225,
+        "eval_runtime": 1103.8475,
+        "eval_samples_per_second": 26.187,
+        "eval_steps_per_second": 1.637,
+        "epoch": 1.1094710947109472,
+        "step": 2255
+    },
+    {
+        "loss": 1.7622,
+        "grad_norm": 0.6831102967262268,
+        "learning_rate": 3.3539145468016795e-05,
+        "epoch": 1.3313653136531365,
+        "step": 2706
+    },
+    {
+        "eval_loss": 1.8246678113937378,
+        "eval_runtime": 1103.4364,
+        "eval_samples_per_second": 26.197,
+        "eval_steps_per_second": 1.638,
+        "epoch": 1.3313653136531365,
+        "step": 2706
+    },
+    {
+        "loss": 1.7536,
+        "grad_norm": 0.6920585036277771,
+        "learning_rate": 2.240059273894789e-05,
+        "epoch": 1.5532595325953258,
+        "step": 3157
+    },
+    {
+        "eval_loss": 1.8157387971878052,
+        "eval_runtime": 1103.7228,
+        "eval_samples_per_second": 26.19,
+        "eval_steps_per_second": 1.637,
+        "epoch": 1.5532595325953258,
+        "step": 3157
+    },
+    {
+        "loss": 1.7467,
+        "grad_norm": 0.6837635040283203,
+        "learning_rate": 1.1262040009878982e-05,
+        "epoch": 1.7751537515375153,
+        "step": 3608
+    },
+    {
+        "eval_loss": 1.808944821357727,
+        "eval_runtime": 1103.6481,
+        "eval_samples_per_second": 26.192,
+        "eval_steps_per_second": 1.637,
+        "epoch": 1.7751537515375153,
+        "step": 3608
+    },
+    {
+        "loss": 1.7457,
+        "grad_norm": 0.6841686367988586,
+        "learning_rate": 1.2348728081007656e-07,
+        "epoch": 1.9970479704797048,
+        "step": 4059
+    },
+    {
+        "eval_loss": 1.805881142616272,
+        "eval_runtime": 1103.2943,
+        "eval_samples_per_second": 26.201,
+        "eval_steps_per_second": 1.638,
+        "epoch": 1.9970479704797048,
+        "step": 4059
+    },
+    {
+        "train_runtime": 69487.1891,
+        "train_samples_per_second": 7.488,
+        "train_steps_per_second": 0.058,
+        "total_flos": 1.0874326325169095e+19,
+        "train_loss": 1.841461892437747,
+        "epoch": 1.9995079950799508,
+        "step": 4064
+    }
+]