Upload vast_zerosec_v3.sh with huggingface_hub

vast_zerosec_v3.sh

@@ -0,0 +1,94 @@
+#!/bin/bash
+set -e
+echo "=== ZeroSec v3 Training (Unsloth QLoRA) ==="
+
+# Install
+pip install unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git -q 2>&1 | tail -1
+pip install "unsloth[cu121-torch250] @ git+https://github.com/unslothai/unsloth.git" -q 2>&1 | tail -1
+pip install torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121 -q 2>&1 | tail -1
+
+# Download datasets
+echo "Downloading datasets..."
+export HF_TOKEN="${HF_TOKEN:?Set HF_TOKEN}"
+curl -L -o exploit_data.jsonl -H "Authorization: Bearer $HF_TOKEN" "https://huggingface.co/HeeBive/ZeroSec-7B/resolve/main/training_data.jsonl"
+curl -L -o tool_data.jsonl -H "Authorization: Bearer $HF_TOKEN" "https://huggingface.co/HeeBive/ZeroSec-7B/resolve/main/v3_tool_data.jsonl"
+
+# Combine
+python3 << 'PYEOF' > train.log 2>&1
+import json, os, torch
+os.environ["TORCH_COMPILE_DISABLE"] = "1"
+HF_TOKEN = os.environ["HF_TOKEN"]
+
+print("Combining datasets...")
+with open("exploit_data.jsonl") as f: exp = f.readlines()
+with open("tool_data.jsonl") as f: tool = f.readlines()
+all_data = exp + tool * 5
+with open("combined.jsonl", "w") as f:
+    for l in all_data: f.write(l)
+print(f"{len(exp)} exploit + {len(tool)*5} tool = {len(all_data)}")
+
+from datasets import load_dataset
+from unsloth import FastLanguageModel
+from trl import SFTTrainer
+from transformers import TrainingArguments
+from unsloth import is_bfloat16_supported
+
+MODEL = "unsloth/gpt-oss-20b-bf16"
+OUT = "/root/zerosec_v3"
+
+print("Loading model...")
+model, tokenizer = FastLanguageModel.from_pretrained(
+    model_name=MODEL,
+    max_seq_length=512,
+    dtype=None,
+    load_in_4bit=True,
+)
+model = FastLanguageModel.get_peft_model(
+    model,
+    r=16,
+    target_modules=["q_proj","k_proj","v_proj","o_proj"],
+    lora_alpha=32,
+    use_gradient_checkpointing="unsloth",
+)
+
+print("Loading dataset...")
+ds = load_dataset("json", data_files="combined.jsonl", split="train")
+def fmt(x): return {"text": x["text"]}
+ds = ds.map(fmt, remove_columns=ds.column_names)
+
+print("Training...")
+trainer = SFTTrainer(
+    model=model,
+    tokenizer=tokenizer,
+    train_dataset=ds,
+    dataset_text_field="text",
+    max_seq_length=512,
+    args=TrainingArguments(
+        per_device_train_batch_size=2,
+        gradient_accumulation_steps=8,
+        num_train_epochs=1,
+        learning_rate=2e-4,
+        fp16=not is_bfloat16_supported(),
+        bf16=is_bfloat16_supported(),
+        logging_steps=10,
+        save_steps=500,
+        output_dir=OUT,
+        optim="adamw_8bit",
+        report_to="none",
+    ),
+)
+trainer.train()
+
+print("Saving...")
+model.save_pretrained(OUT)
+tokenizer.save_pretrained(OUT)
+
+print("Uploading to HF...")
+from huggingface_hub import HfApi
+api = HfApi(token=HF_TOKEN)
+api.create_repo("HeeBive/ZeroSec-v3", repo_type="model", exist_ok=True, private=True)
+api.upload_folder(folder_path=OUT, repo_id="HeeBive/ZeroSec-v3", repo_type="model", token=HF_TOKEN)
+print("✅ DONE - HeeBive/ZeroSec-v3")
+PYEOF
+
+tail -f train.log