Upload kassandra-lora.py

example/kassandra-lora.py

@@ -0,0 +1,131 @@
+import torch
+import random
+import numpy as np
+import gc
+from datasets import load_dataset
+from transformers import (
+    AutoTokenizer,
+    AutoModelForCausalLM,
+    Trainer,
+    TrainingArguments,
+    DataCollatorForLanguageModeling,
+)
+from peft import LoraConfig, get_peft_model
+
+# ============================================================
+# Seed
+# ============================================================
+def set_seed(seed=42):
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)
+
+set_seed(42)
+
+# ============================================================
+# Model & Tokenizer
+# ============================================================
+MODEL_ID = "/opt/models/mistralai/Mistral-7B-Instruct-v0.3"
+DATA_PATH = "sebelsn/style-adjustment-dataset_de/2026-02-06_style-adjustment-dataset_de.jsonl"
+
+tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, use_fast=True)
+if tokenizer.pad_token is None:
+    tokenizer.pad_token = tokenizer.eos_token
+
+model = AutoModelForCausalLM.from_pretrained(
+    MODEL_ID,
+    dtype=torch.bfloat16,
+    device_map="cuda",
+)
+model.config.pad_token_id = tokenizer.pad_token_id  # ← NEU
+
+# Optional: Gradient Checkpointing
+# model.gradient_checkpointing_enable()
+
+# ============================================================
+# LoRA
+# ============================================================
+lora_config = LoraConfig(
+    r=1,
+    lora_alpha=2,
+    target_modules=["q_proj", "v_proj"],
+    lora_dropout=0.1,
+    bias="none",
+    task_type="CAUSAL_LM",
+)
+model = get_peft_model(model, lora_config)
+model.print_trainable_parameters()
+
+# ============================================================
+# Dataset
+# ============================================================
+dataset = load_dataset("json", data_files=DATA_PATH)["train"]
+dataset = dataset.shuffle(seed=42)  # ← NEU
+
+def format_example(example):
+    text = f"Frage:\n{example['instruction']}\n\nAntwort:\n{example['response']}"
+    return {"text": text}
+
+dataset = dataset.map(format_example, remove_columns=dataset.column_names)
+
+def tokenize(example):
+    return tokenizer(example["text"], truncation=True, max_length=512, padding=False)
+
+dataset = dataset.map(tokenize, batched=True)
+
+# Train/Val Split
+dataset = dataset.train_test_split(test_size=0.1, seed=42)  # ← NEU
+train_dataset = dataset["train"]
+val_dataset = dataset["test"]
+
+# ============================================================
+# Training
+# ============================================================
+data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
+
+training_args = TrainingArguments(
+    output_dir="/opt/models/lora-style-out",
+    per_device_train_batch_size=1,
+    gradient_accumulation_steps=8,
+    learning_rate=1e-5,                # ← Etwas höher
+    lr_scheduler_type="cosine",        # ← NEU
+    warmup_ratio=0.5,                  # ← Mehr warmup
+    max_grad_norm=1.0,                 # ← NEU
+    num_train_epochs=2.5,
+    bf16=True,
+    max_steps=56,
+    logging_steps=7,
+    logging_dir="/opt/models/lora-style-out/logs",  # ← NEU
+    save_strategy="steps",
+    save_steps=14,
+    eval_strategy="steps",             # ← NEU
+    eval_steps=5,                      # ← NEU
+    load_best_model_at_end=False,      # ← NEU
+    metric_for_best_model="loss",      # ← NEU
+    report_to="none",
+)
+
+trainer = Trainer(
+    model=model,
+    args=training_args,
+    train_dataset=train_dataset,       # ← Geändert
+    eval_dataset=val_dataset,          # ← NEU
+    data_collator=data_collator,
+)
+
+trainer.train()
+
+# ============================================================
+# Save & Cleanup
+# ============================================================
+#trainer.save_model("/opt/models/lora-style-out/final")  # ← NEU
+#tokenizer.save_pretrained("/opt/models/lora-style-out/final")
+
+del model
+del trainer
+torch.cuda.empty_cache()
+gc.collect()
+torch.cuda.synchronize()
+
+print("✅ Training done, GPU freed")