feat: QLoRA fine-tune v1 on M5 + fused model (M2)

configs/lora.yaml

@@ -0,0 +1,16 @@
+model: mlx-community/Qwen2.5-7B-Instruct-4bit
+train: true
+data: data/train
+adapter_path: adapters/buddhagpt-v1
+batch_size: 1
+grad_accumulation_steps: 8
+iters: 1200
+learning_rate: 1e-5
+num_layers: 16
+lora_parameters:
+  rank: 16
+  scale: 20.0
+  dropout: 0.05
+max_seq_length: 1024
+steps_per_eval: 200
+save_every: 200

scripts/prepare_train.py

@@ -0,0 +1,12 @@
+import json, random
+from pathlib import Path
+
+rows = [json.loads(l) for l in Path("data/instructions.jsonl").read_text().splitlines()]
+random.seed(7); random.shuffle(rows)
+n_val = max(200, len(rows) // 20)
+Path("data/train").mkdir(parents=True, exist_ok=True)
+for name, part in [("valid", rows[:n_val]), ("train", rows[n_val:])]:
+    with Path(f"data/train/{name}.jsonl").open("w") as f:
+        for r in part:
+            f.write(json.dumps(r) + "\n")
+print(len(rows) - n_val, "train /", n_val, "valid")