Create app.py

app.py

@@ -0,0 +1,190 @@
+import os
+import json
+from pathlib import Path
+from datetime import datetime
+
+from datasets import load_dataset, DatasetDict
+from huggingface_hub import login, create_repo, upload_folder
+from transformers import (
+    AutoTokenizer,
+    AutoModelForCausalLM,
+    DataCollatorForSeq2Seq,
+    TrainingArguments,
+    Trainer,
+)
+from peft import LoraConfig, get_peft_model
+
+# ------------- Config utilisateur -------------
+HF_TOKEN = os.environ["HF_TOKEN"]  # ⚠️ récupère ton token depuis l'environnement
+BASE_MODEL_ID = "Gopu-poss/gopu-agent-2k-fdf"
+ALIGNED_MODEL_ID = "Gopu-poss/gopu-agent-2k-fdf-aligned"
+DATASET_REPO_ID = "Gopu-poss/gopu-agent-2k-fdf-dataset-prepared"
+OUTPUT_DIR = "./gopu-lora-out"
+SEED = 42
+
+# ------------- Auth -------------
+print(">> Logging into Hugging Face Hub...")
+login(token=HF_TOKEN)
+
+# ------------- Chargement des datasets bruts -------------
+print(">> Loading HuggingFaceFW/finewiki (train split)...")
+finewiki = load_dataset("HuggingFaceFW/finewiki", split="train")
+print(">> Loading fka/awesome-chatgpt-prompts (train split)...")
+awesome = load_dataset("fka/awesome-chatgpt-prompts", split="train")
+
+# ------------- Échantillonnage / préparation -------------
+FW_SAMPLE_SIZE = 20000
+if len(finewiki) > FW_SAMPLE_SIZE:
+    finewiki = finewiki.shuffle(seed=SEED).select(range(FW_SAMPLE_SIZE))
+
+print(f">> finewiki sampled: {len(finewiki)} rows; awesome: {len(awesome)} rows")
+
+# ------------- Normalisation en instruction / input / output -------------
+def map_finewiki(example):
+    title = example.get("title", "")
+    text = example.get("text", "")
+    instruction = f"Explique en termes clairs et techniques l'article: {title}"
+    input_ctx = text[:2000]
+    output = (
+        "Résumé technique et stylisé (GopuOS): "
+        "Points clés, concepts, et relations. Maintiens un ton clair, concis, et agentique."
+    )
+    return {"instruction": instruction, "input": input_ctx, "output": output}
+
+finewiki_mapped = finewiki.map(map_finewiki)
+
+def map_awesome(example):
+    act = example.get("act", "")
+    prompt = example.get("prompt", "")
+    instruction = f"Rôle/acte: {act}. Réponds au prompt en style GopuOS."
+    input_ctx = prompt
+    output = (
+        "Réponse alignée GopuOS: concise, technique, introspectable, bilingue possible FR/EN."
+    )
+    return {"instruction": instruction, "input": input_ctx, "output": output}
+
+awesome_mapped = awesome.map(map_awesome)
+
+prepared = DatasetDict({
+    "train": finewiki_mapped,
+    "eval": awesome_mapped
+})
+
+# ------------- Sauvegarde locale du dataset prétraité -------------
+prepared_dir = Path("./prepared_dataset")
+prepared_dir.mkdir(parents=True, exist_ok=True)
+for split in prepared.keys():
+    out_path = prepared_dir / f"{split}.jsonl"
+    with out_path.open("w", encoding="utf-8") as f:
+        for ex in prepared[split]:
+            f.write(json.dumps(ex, ensure_ascii=False) + "\n")
+
+# ------------- Push du dataset prétraité sur le Hub -------------
+print(f">> Creating/updating dataset repo: {DATASET_REPO_ID}")
+create_repo(repo_id=DATASET_REPO_ID, token=HF_TOKEN, repo_type="dataset", private=False, exist_ok=True)
+
+upload_folder(
+    repo_id=DATASET_REPO_ID,
+    repo_type="dataset",
+    folder_path=str(prepared_dir),
+    token=HF_TOKEN,
+    commit_message=f"Prepared dataset push {datetime.utcnow().isoformat()}",
+)
+
+# ------------- Chargement modèle/tokenizer -------------
+print(f">> Loading base model: {BASE_MODEL_ID}")
+tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL_ID, use_auth_token=HF_TOKEN)
+model = AutoModelForCausalLM.from_pretrained(BASE_MODEL_ID, use_auth_token=HF_TOKEN)
+
+# ------------- PEFT LoRA config -------------
+peft_config = LoraConfig(
+    r=8,
+    lora_alpha=16,
+    target_modules=["q_proj", "v_proj"],
+    lora_dropout=0.05,
+    bias="none",
+    task_type="CAUSAL_LM",
+)
+model = get_peft_model(model, peft_config)
+
+# ------------- Tokenization -------------
+def format_example(ex):
+    system_prompt = (
+        "Tu es Gopu, agent intelligent de GopuOS. Réponds de manière claire, technique, stylisée, et introspectable."
+    )
+    user = f"Utilisateur: {ex['instruction']}\nContexte: {ex['input']}\nGopu:"
+    target = ex["output"]
+    src = system_prompt + "\n\n" + user
+    return {"src": src, "tgt": target}
+
+formatted = prepared.map(format_example)
+
+def tokenize(batch):
+    model_inputs = tokenizer(
+        batch["src"],
+        truncation=True,
+        max_length=1024,
+    )
+    with tokenizer.as_target_tokenizer():
+        labels = tokenizer(
+            batch["tgt"],
+            truncation=True,
+            max_length=256,
+        )
+    model_inputs["labels"] = labels["input_ids"]
+    return model_inputs
+
+tokenized_train = formatted["train"].map(tokenize, batched=False, remove_columns=formatted["train"].column_names)
+tokenized_eval = formatted["eval"].map(tokenize, batched=False, remove_columns=formatted["eval"].column_names)
+
+data_collator = DataCollatorForSeq2Seq(tokenizer=tokenizer, model=model, padding="longest")
+
+# ------------- Entraînement -------------
+training_args = TrainingArguments(
+    output_dir=OUTPUT_DIR,
+    per_device_train_batch_size=4,
+    per_device_eval_batch_size=4,
+    gradient_accumulation_steps=2,
+    eval_strategy="steps",
+    eval_steps=200,
+    logging_steps=50,
+    save_steps=500,
+    save_total_limit=2,
+    num_train_epochs=1,
+    learning_rate=2e-4,
+    warmup_steps=200,
+    weight_decay=0.01,
+    fp16=True,
+    bf16=False,
+    report_to=[],
+    seed=SEED,
+)
+
+trainer = Trainer(
+    model=model,
+    args=training_args,
+    data_collator=data_collator,
+    train_dataset=tokenized_train,
+    eval_dataset=tokenized_eval,
+)
+
+print(">> Starting training...")
+trainer.train()
+print(">> Training complete")
+
+# ------------- Sauvegarde et push du modèle -------------
+print(f">> Creating/updating model repo: {ALIGNED_MODEL_ID}")
+create_repo(repo_id=ALIGNED_MODEL_ID, token=HF_TOKEN, repo_type="model", private=False, exist_ok=True)
+
+trainer.save_model(OUTPUT_DIR)
+tokenizer.save_pretrained(OUTPUT_DIR)
+
+upload_folder(
+    repo_id=ALIGNED_MODEL_ID,
+    repo_type="model",
+    folder_path=OUTPUT_DIR,
+    token=HF_TOKEN,
+    commit_message=f"Push aligned LoRA model {datetime.utcnow().isoformat()}",
+)
+
+print(f">> Model pushed: https://huggingface.co/{ALIGNED_MODEL_ID}")