Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

TrafficoDataset/train_gemma3_traffico.py

@@ -0,0 +1,241 @@
+# ============================================================
+# Gemma-3-270M – Analisi Traffico di Rete TCP/IP
+# Unsloth + LoRA + Dataset JSONL (CIC-IDS2017 + UNSW-NB15)
+# ============================================================
+# Struttura basata sul tuo script, adattata per il dominio
+# di analisi del traffico di rete con mappatura MITRE ATT&CK.
+#
+# PREREQUISITI:
+#   Google Colab con runtime GPU (T4 basta)
+#   !pip install --no-deps unsloth
+#   !pip install transformers datasets trl peft accelerate sentencepiece
+#
+# FILE NECESSARI (nella stessa cartella dello script):
+#   dataset.jsonl   ← generato dalla script apposita
+# ============================================================
+
+# ---------- INSTALL (Colab) ----------
+# !pip install --no-deps unsloth
+# !pip install transformers datasets trl peft accelerate sentencepiece
+
+# ---------- IMPORT ----------
+from unsloth import FastModel
+from unsloth.chat_templates import get_chat_template, train_on_responses_only
+import torch
+from datasets import load_dataset
+from trl import SFTTrainer, SFTConfig
+
+# ---------- CONFIG ----------
+MODEL_NAME   = "unsloth/gemma-3-270m-it"
+DATASET_PATH = "dataset_traffico.jsonl"          # <== il JSONL che abbiamo generato
+OUTPUT_DIR   = "outputs"
+MAX_SEQ_LENGTH = 512                    # 512 basta per questi prompt, risparmia memoria
+
+# ---------- LOAD MODEL ----------
+model, tokenizer = FastModel.from_pretrained(
+    model_name     = MODEL_NAME,
+    max_seq_length = MAX_SEQ_LENGTH,
+    load_in_4bit   = False,
+    load_in_8bit   = False,
+    full_finetuning = False,
+)
+
+# ---------- LoRA ----------
+# Configurazione più aggressiva sul rank (r=64) per un dominio specifico come questo.
+# Target modules: tutti i proiettori del transformer.
+model = FastModel.get_peft_model(
+    model,
+    r              = 64,
+    target_modules = [
+        "q_proj", "k_proj", "v_proj", "o_proj",
+        "gate_proj", "up_proj", "down_proj",
+    ],
+    lora_alpha     = 64,
+    lora_dropout   = 0,
+    bias           = "none",
+    use_gradient_checkpointing = "unsloth",
+    random_state   = 3407,
+)
+
+# ---------- CHAT TEMPLATE (Gemma-3) ----------
+tokenizer = get_chat_template(
+    tokenizer,
+    chat_template = "gemma3",
+)
+
+# ---------- LOAD DATASET ----------
+dataset = load_dataset(
+    "json",
+    data_files = DATASET_PATH,
+    split      = "train",
+)
+
+print(f"Dataset caricato: {len(dataset)} righe")
+print(f"Campi presenti: {dataset.column_names}")
+print(f"\nEsempio riga 0:")
+print(dataset[0])
+
+# ---------- CONVERT TO CHATML ----------
+# Il JSONL ha campi: instruction, input, output
+# Li convertiamo nel formato conversations [system, user, assistant]
+# che Gemma-3 si aspetta.
+def convert_to_chatml(example):
+    system_prompt = example["instruction"]
+
+    # Se c'è un campo 'context' lo aggiungiamo al system prompt
+    if "context" in example and example["context"]:
+        system_prompt += f"\nContesto: {example['context']}."
+
+    return {
+        "conversations": [
+            {"role": "system",    "content": system_prompt},
+            {"role": "user",      "content": example["input"]},
+            {"role": "assistant", "content": example["output"]},
+        ]
+    }
+
+dataset = dataset.map(convert_to_chatml)
+
+# ---------- APPLY GEMMA-3 TEMPLATE ----------
+# Applica il template di chat di Gemma-3 a ogni esempio.
+# Questo produce la stringa finale che il modello vedrà durante il training.
+def formatting_prompts_func(examples):
+    convos = examples["conversations"]
+    texts = [
+        tokenizer.apply_chat_template(
+            convo,
+            tokenize            = False,
+            add_generation_prompt = False,
+        ).removeprefix("<bos>")
+        for convo in convos
+    ]
+    return {"text": texts}
+
+dataset = dataset.map(formatting_prompts_func, batched=True)
+
+# Verifica come appare un prompt formattato
+print("\n" + "=" * 60)
+print(" PROMPT FORMATTATO (esempio)")
+print("=" * 60)
+print(dataset[0]["text"])
+print("=" * 60)
+
+# ---------- TRAINER ----------
+trainer = SFTTrainer(
+    model          = model,
+    tokenizer      = tokenizer,
+    train_dataset  = dataset,
+    eval_dataset   = None,
+    args = SFTConfig(
+        dataset_text_field          = "text",
+        per_device_train_batch_size = 4,
+        gradient_accumulation_steps = 4,       # batch effettivo = 4 * 4 = 16
+        warmup_steps                = 10,
+        max_steps                   = 500,     # ~500 step su 10k righe con batch 16
+        learning_rate               = 2e-5,
+        logging_steps               = 25,
+        optim                       = "adamw_8bit",
+        weight_decay                = 0.001,
+        lr_scheduler_type           = "linear",
+        seed                        = 3407,
+        output_dir                  = OUTPUT_DIR,
+        report_to                   = "none",
+    ),
+)
+
+# ---------- TRAIN ONLY ON ASSISTANT ----------
+# Fondamentale: il modello calcola il loss SOLO sulla risposta dell'assistant,
+# non sul prompt. Così non "impara" a ripetere la domanda.
+trainer = train_on_responses_only(
+    trainer,
+    instruction_part = "<start_of_turn>user\n",
+    response_part    = "<start_of_turn>model\n",
+)
+
+# ---------- TRAIN ----------
+trainer.train()
+
+# ---------- SAVE LoRA ----------
+model.save_pretrained("gemma3-traffico-rete-lora")
+tokenizer.save_pretrained("gemma3-traffico-rete-lora")
+print("\n✓ Modello LoRA salvato in: gemma3-traffico-rete-lora/")
+model.save_pretrained_merged(
+    "gemma3-traffico-rete-lora",  # cartella output
+    tokenizer,
+    save_method="merged_16bit"       # Float16 per GGUF
+)
+model.save_pretrained_gguf(
+        "gemma3-traffico-rete-lora",
+        tokenizer,
+        quantization_method = "BF16", # For now only Q8_0, BF16, F16 supported
+    )
+# ---------- INFERENCE: TEST ----------
+# Dopo il training, prova il modello con alcuni flussi di esempio.
+from transformers import TextStreamer
+
+test_cases = [
+    # Caso 1: profilo tipico DoS (masse enormi di byte src, pochissimi dst, durata minima)
+    "Protocollo: tcp | Porta dst: 80 | Byte src: 480000 | Byte dst: 40 | Pacchetti: 5200 | Durata: 0.015s",
+    # Caso 2: traffico normale HTTPS
+    "Protocollo: tcp | Porta dst: 443 | Byte src: 1500 | Byte dst: 6200 | Pacchetti: 9 | Durata: 3.200s",
+    # Caso 3: profilo PortScan (tanti dst diversi, pochi byte, durata quasi zero)
+    "Protocollo: tcp | Porta dst: 22 | Byte src: 60 | Byte dst: 0 | Pacchetti: 1 | Durata: 0.002s",
+    # Caso 4: profilo Brute Force su SSH
+    "Protocollo: tcp | Porta dst: 22 | Byte src: 3200 | Byte dst: 8500 | Pacchetti: 45 | Durata: 1.800s",
+    # Caso 5: profilo Infiltration / esfiltrazioni dati
+    "Protocollo: tcp | Porta dst: 443 | Byte src: 8000 | Byte dst: 120000 | Pacchetti: 200 | Durata: 25.500s",
+]
+
+streamer = TextStreamer(tokenizer, skip_prompt=True)
+
+for i, test_input in enumerate(test_cases, 1):
+    messages = [
+        {
+            "role": "system",
+            "content": (
+                "Analizza il seguente flusso di traffico di rete TCP/IP. "
+                "Classifica se è traffico normale o un attacco. "
+                "Se è un attacco, indica la categoria e la tecnica MITRE ATT&CK corrispondente."
+            ),
+        },
+        {"role": "user", "content": test_input},
+    ]
+
+    text = tokenizer.apply_chat_template(
+        messages,
+        tokenize            = False,
+        add_generation_prompt = True,
+    ).removeprefix("<bos>")
+
+    print(f"\n{'─' * 60}")
+    print(f" TEST {i}: {test_input[:80]}...")
+    print(f"{'─' * 60}")
+    print(" Risposta: ", end="")
+
+    _ = model.generate(
+        **tokenizer(text, return_tensors="pt").to("cuda"),
+        max_new_tokens = 128,
+        temperature    = 0.3,       # bassa temperatura = risposte più deterministe
+        top_p          = 0.9,
+        top_k          = 40,
+        streamer       = streamer,
+    )
+
+# ---------- SAVE MERGED (opzionale) ----------
+# Unisce i pesi LoRA al modello base e salva come modello completo.
+# Utile per deployare senza dipendenza da PEFT.
+#
+model.save_pretrained_merged(
+    "gemma3-traffico-rete-merged",
+    tokenizer,
+    save_method = "merged_16bit",
+)
+#
+# ---------- SAVE GGUF (opzionale) ----------
+# Formato GGUF per inferenza locale con llama.cpp / Ollama.
+#
+model.save_pretrained_gguf(
+    "gemma3-traffico-rete-gguf",
+    tokenizer,
+    quantization_method = "Q8_0",   # Q8_0 = buon equilibrio qualità/dimensione
+)

added_tokens.json

@@ -0,0 +1,3 @@
+{
+  "<image_soft_token>": 262144
+}

chat_template.jinja

@@ -0,0 +1,47 @@
+
+{%- if messages[0]['role'] == 'system' -%}
+    {%- if messages[0]['content'] is string -%}
+        {%- set first_user_prefix = messages[0]['content'] + '
+
+' -%}
+    {%- else -%}
+        {%- set first_user_prefix = messages[0]['content'][0]['text'] + '
+
+' -%}
+    {%- endif -%}
+    {%- set loop_messages = messages[1:] -%}
+{%- else -%}
+    {%- set first_user_prefix = "" -%}
+    {%- set loop_messages = messages -%}
+{%- endif -%}
+{%- for message in loop_messages -%}
+    {%- if (message['role'] == 'user') != (loop.index0 % 2 == 0) -%}
+        {{ raise_exception("Conversation roles must alternate user/assistant/user/assistant/...") }}
+    {%- endif -%}
+    {%- if (message['role'] == 'assistant') -%}
+        {%- set role = "model" -%}
+    {%- else -%}
+        {%- set role = message['role'] -%}
+    {%- endif -%}
+    {{ '<start_of_turn>' + role + '
+' + (first_user_prefix if loop.first else "") }}
+    {%- if message['content'] is string -%}
+        {{ message['content'] | trim }}
+    {%- elif message['content'] is iterable -%}
+        {%- for item in message['content'] -%}
+            {%- if item['type'] == 'image' -%}
+                {{ '<start_of_image>' }}
+            {%- elif item['type'] == 'text' -%}
+                {{ item['text'] | trim }}
+            {%- endif -%}
+        {%- endfor -%}
+    {%- else -%}
+        {{ raise_exception("Invalid content type") }}
+    {%- endif -%}
+    {{ '<end_of_turn>
+' }}
+{%- endfor -%}
+{%- if add_generation_prompt -%}
+    {{ '<start_of_turn>model
+' }}
+{%- endif -%}

config.json

@@ -0,0 +1,56 @@
+{
+  "_sliding_window_pattern": 6,
+  "architectures": [
+    "Gemma3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "attn_logit_softcapping": null,
+  "bos_token_id": 2,
+  "torch_dtype": "bfloat16",
+  "eos_token_id": 106,
+  "final_logit_softcapping": null,
+  "head_dim": 256,
+  "hidden_activation": "gelu_pytorch_tanh",
+  "hidden_size": 640,
+  "initializer_range": 0.02,
+  "intermediate_size": 2048,
+  "layer_types": [
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 32768,
+  "model_type": "gemma3_text",
+  "num_attention_heads": 4,
+  "num_hidden_layers": 18,
+  "num_key_value_heads": 1,
+  "pad_token_id": 0,
+  "query_pre_attn_scalar": 256,
+  "rms_norm_eps": 1e-06,
+  "rope_local_base_freq": 10000.0,
+  "rope_scaling": null,
+  "rope_theta": 1000000.0,
+  "sliding_window": 512,
+  "transformers_version": "4.57.3",
+  "unsloth_fixed": true,
+  "unsloth_version": "2026.1.4",
+  "use_bidirectional_attention": false,
+  "use_cache": true,
+  "vocab_size": 262144
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1debdbdbc6a711e1abf5f0285bda7fd2a7a93805ae3c4aa986012a7bd2eac39a
+size 536223056

special_tokens_map.json

@@ -0,0 +1,33 @@
+{
+  "boi_token": "<start_of_image>",
+  "bos_token": {
+    "content": "<bos>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eoi_token": "<end_of_image>",
+  "eos_token": {
+    "content": "<end_of_turn>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "image_token": "<image_soft_token>",
+  "pad_token": {
+    "content": "<pad>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4667f2089529e8e7657cfb6d1c19910ae71ff5f28aa7ab2ff2763330affad795
+size 33384568

tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1299c11d7cf632ef3b4e11937501358ada021bbdf7c47638d13c0ee982f2e79c
+size 4689074