chore: add .gitignore

runpod/train_runpod.py

@@ -2,17 +2,15 @@
 """
 Fine-tuning QLoRA de gemma3:12b sur RunPod (A100 40/80GB).
 
-Différences vs local (RTX 5070 12GB) :
-  - max_seq_length=2048 (vs 512) — prompts pipeline non tronqués
-  - batch_size=4 (vs 1) — convergence plus stable
-  - 3 epochs (vs 1) — meilleure mémorisation
-  - Dataset complet sans sous-échantillonnage
+Dataset V2 : 15.9K train, 53% raisonnement structuré (vs 95% lookups V1).
+Sources : referentiels, pipeline, cocoa, ccam, cim10, reasoning,
+          negative, discrimination, fascicule_reasoning, guide_metho.
 
 Usage sur RunPod :
-  1. Créer un pod avec template PyTorch 2.4+ / CUDA 12.x
-  2. rsync -avz runpod/ runpod_host:/workspace/t2a-finetune/
-  3. bash /workspace/t2a-finetune/setup.sh
-  4. python /workspace/t2a-finetune/train_runpod.py [--epochs 3] [--export-gguf]
+  1. Créer un pod A100 80GB (template PyTorch 2.4+ / CUDA 12.x)
+  2. Uploader les fichiers (train_runpod.py, setup.sh, data/)
+  3. bash setup.sh
+  4. python train_runpod.py [--epochs 3] [--export-gguf]
 """
 
 import argparse
@@ -149,7 +147,7 @@ def train(model, tokenizer, train_ds, eval_ds, args):
     callbacks = []
     try:
         import wandb
-        wandb.init(project="pmsi-coder", name=f"runpod-{args.epochs}ep-seq{args.max_seq_length}")
+        wandb.init(project="pmsi-coder", name=f"v2-runpod-{args.epochs}ep-seq{args.max_seq_length}")
         report = "wandb"
         print("  Tracking : wandb")
     except ImportError:
@@ -205,34 +203,61 @@ def train(model, tokenizer, train_ds, eval_ds, args):
     return trainer, final_dir
 
 
+def export_merged_hf(model, tokenizer):
+    """Sauvegarder le modèle mergé en 16-bit (HF format) pour conversion GGUF ultérieure."""
+    print(f"\nExport modèle mergé (16-bit HF)...")
+
+    merged_dir = OUTPUT / "pmsi-merged-hf"
+    merged_dir.mkdir(parents=True, exist_ok=True)
+
+    model.save_pretrained_merged(
+        str(merged_dir),
+        tokenizer,
+        save_method="merged_16bit",
+    )
+
+    size_gb = sum(f.stat().st_size for f in merged_dir.glob("*.safetensors")) / 1024**3
+    print(f"  Modèle mergé : {merged_dir} ({size_gb:.1f} Go)")
+    print(f"\n  Pour convertir en GGUF :")
+    print(f"    python llama.cpp/convert_hf_to_gguf.py {merged_dir} --outfile pmsi-v2-q8.gguf --outtype q8_0")
+    print(f"    llama-quantize pmsi-v2-q8.gguf pmsi-v2-q4km.gguf Q4_K_M")
+    return merged_dir
+
+
 def export_gguf(model, tokenizer, final_dir, quantization="q4_k_m"):
+    """Export GGUF via Unsloth (peut échouer — fallback sur export_merged_hf)."""
     print(f"\nExport GGUF ({quantization})...")
 
     gguf_dir = OUTPUT / "pmsi-gguf"
     gguf_dir.mkdir(parents=True, exist_ok=True)
 
-    model.save_pretrained_gguf(
-        str(gguf_dir),
-        tokenizer,
-        quantization_method=quantization,
-    )
+    try:
+        model.save_pretrained_gguf(
+            str(gguf_dir),
+            tokenizer,
+            quantization_method=quantization,
+        )
 
-    gguf_files = list(gguf_dir.glob("*.gguf"))
-    if gguf_files:
-        gguf_path = gguf_files[0]
-        size_gb = gguf_path.stat().st_size / 1024**3
-        print(f"  GGUF : {gguf_path.name} ({size_gb:.1f} Go)")
+        gguf_files = list(gguf_dir.glob("*.gguf"))
+        if gguf_files:
+            gguf_path = gguf_files[0]
+            size_gb = gguf_path.stat().st_size / 1024**3
+            print(f"  GGUF : {gguf_path.name} ({size_gb:.1f} Go)")
 
-        modelfile_path = gguf_dir / "Modelfile"
-        with open(modelfile_path, "w") as f:
-            f.write(f"FROM {gguf_path.name}\n\n")
-            f.write("PARAMETER temperature 0.3\n")
-            f.write("PARAMETER top_p 0.9\n")
-            f.write("PARAMETER num_ctx 8192\n")
+            modelfile_path = gguf_dir / "Modelfile"
+            with open(modelfile_path, "w") as f:
+                f.write(f"FROM {gguf_path.name}\n\n")
+                f.write("PARAMETER temperature 0.3\n")
+                f.write("PARAMETER top_p 0.9\n")
+                f.write("PARAMETER num_ctx 8192\n")
+                f.write('PARAMETER stop "<end_of_turn>"\n')
+                f.write('PARAMETER stop "<eos>"\n')
 
-        print(f"  Modelfile créé")
-        print(f"\n  Pour récupérer : scp runpod:{gguf_path} .")
-        print(f"  Puis : ollama create pmsi-coder -f Modelfile")
+            print(f"  Modelfile créé")
+    except Exception as e:
+        print(f"  GGUF export échoué : {e}")
+        print(f"  Fallback : export HF mergé...")
+        export_merged_hf(model, tokenizer)
 
 
 def main():