feat: tts-norm LoRA — dataset generator + training script

gen_tts_dataset.py: 4960 synthetic examples, 22 categories (numbers,
currencies, dates, times, temperatures, acronyms, NATO phonetic, URLs,
markdown, etc). Bilingual EN/PL with explicit [lang] tag prefix.

train_tts_norm.py: Unsloth LoRA training for Qwen2.5-7B-Instruct.
Rank 16, 3 epochs, packing, max_seq 768. Trained on H100 in 20m38s,
final loss 0.091. Adapter: 154MB.

train_tts_norm.py

@@ -0,0 +1,173 @@
+#!/usr/bin/env python3
+"""Train TTS normalization LoRA on Qwen2.5-7B-Instruct using Unsloth.
+
+Reads: tts_norm_dataset.jsonl (ShareGPT format)
+Output: tts-norm-lora/ adapter (vLLM-compatible)
+
+Run on junkpile — RTX 2000 Ada 16GB.
+"""
+
+import os
+import torch
+from pathlib import Path
+from unsloth import FastLanguageModel
+from unsloth.chat_templates import get_chat_template, standardize_sharegpt
+from trl import SFTTrainer
+from transformers import TrainingArguments
+from datasets import load_dataset
+
+# ──────────────────────────────────────────────────────────────
+# CONFIG
+# ──────────────────────────────────────────────────────────────
+
+MODEL_NAME = "unsloth/Qwen2.5-7B-Instruct-bnb-4bit"
+DATASET_PATH = "tts_norm_dataset.jsonl"
+OUTPUT_DIR = "./tts-norm-lora"
+MAX_SEQ_LEN = 768         # TTS normalization is short text
+LORA_RANK = 16
+LORA_ALPHA = 16
+BATCH_SIZE = 2             # small GPU — use grad accumulation
+GRAD_ACCUM = 8             # effective batch = 16
+EPOCHS = 3
+LR = 2e-4
+WARMUP_STEPS = 30
+SAVE_STEPS = 100
+LOGGING_STEPS = 10
+SEED = 42
+
+# ──────────────────────────────────────────────────────────────
+# LOAD MODEL
+# ──────────────────────────────────────────────────────────────
+
+print(f"Loading {MODEL_NAME}...")
+model, tokenizer = FastLanguageModel.from_pretrained(
+    model_name=MODEL_NAME,
+    max_seq_length=MAX_SEQ_LEN,
+    load_in_4bit=True,
+    dtype=None,  # auto-detect
+)
+
+# Apply chat template
+tokenizer = get_chat_template(
+    tokenizer,
+    chat_template="qwen-2.5",
+)
+
+# ──────────────────────────────────────────────────────────────
+# PEFT CONFIG
+# ──────────────────────────────────────────────────────────────
+
+print("Applying LoRA...")
+model = FastLanguageModel.get_peft_model(
+    model,
+    r=LORA_RANK,
+    lora_alpha=LORA_ALPHA,
+    target_modules=[
+        "q_proj", "k_proj", "v_proj", "o_proj",
+        "gate_proj", "up_proj", "down_proj",
+    ],
+    lora_dropout=0,
+    bias="none",
+    use_gradient_checkpointing="unsloth",
+    random_state=SEED,
+)
+
+# ──────────────────────────────────────────────────────────────
+# DATASET
+# ──────────────────────────────────────────────────────────────
+
+print(f"Loading dataset from {DATASET_PATH}...")
+dataset = load_dataset("json", data_files=DATASET_PATH, split="train")
+print(f"  {len(dataset)} examples loaded")
+
+# Standardize to ShareGPT format (handles from/value vs role/content)
+dataset = standardize_sharegpt(dataset)
+
+# Pre-apply chat template via map — avoids formatting_func signature issues
+def apply_template(examples):
+    """Apply Qwen2.5 chat template to conversations."""
+    convos = examples["conversations"]
+    texts = []
+    for convo in convos:
+        text = tokenizer.apply_chat_template(
+            convo,
+            tokenize=False,
+            add_generation_prompt=False,
+        )
+        texts.append(text)
+    return {"text": texts}
+
+print("Applying chat template...")
+dataset = dataset.map(apply_template, batched=True, num_proc=2)
+
+# ──────────────────────────────────────────────────────────────
+# TRAINER
+# ──────────────────────────────────────────────────────────────
+
+print("Setting up trainer...")
+trainer = SFTTrainer(
+    model=model,
+    tokenizer=tokenizer,
+    train_dataset=dataset,
+    dataset_text_field="text",
+    args=TrainingArguments(
+        output_dir=OUTPUT_DIR,
+        per_device_train_batch_size=BATCH_SIZE,
+        gradient_accumulation_steps=GRAD_ACCUM,
+        num_train_epochs=EPOCHS,
+        learning_rate=LR,
+        lr_scheduler_type="cosine",
+        warmup_steps=WARMUP_STEPS,
+        fp16=not torch.cuda.is_bf16_supported(),
+        bf16=torch.cuda.is_bf16_supported(),
+        logging_steps=LOGGING_STEPS,
+        save_steps=SAVE_STEPS,
+        save_total_limit=3,
+        seed=SEED,
+        optim="adamw_8bit",
+        weight_decay=0.01,
+        max_grad_norm=1.0,
+        report_to="none",
+        dataloader_num_workers=2,
+    ),
+    max_seq_length=MAX_SEQ_LEN,
+    dataset_num_proc=2,
+    packing=True,  # pack short examples for efficiency
+)
+
+# ──────────────────────────────────────────────────────────────
+# TRAIN
+# ──────────────────────────────────────────────────────────────
+
+print("Starting training...")
+stats = trainer.train()
+print(f"\nTraining complete!")
+print(f"  Total steps: {stats.global_step}")
+print(f"  Train loss: {stats.training_loss:.4f}")
+print(f"  Runtime: {stats.metrics['train_runtime']:.0f}s")
+
+# ──────────────────────────────────────────────────────────────
+# SAVE ADAPTER
+# ──────────────────────────────────────────────────────────────
+
+print(f"\nSaving adapter to {OUTPUT_DIR}...")
+model.save_pretrained(OUTPUT_DIR)
+tokenizer.save_pretrained(OUTPUT_DIR)
+
+# Verify
+adapter_path = Path(OUTPUT_DIR) / "adapter_model.safetensors"
+if adapter_path.exists():
+    size_mb = adapter_path.stat().st_size / (1024 * 1024)
+    print(f"  Adapter saved: {size_mb:.1f} MB")
+else:
+    print("  WARNING: adapter_model.safetensors not found!")
+
+config_path = Path(OUTPUT_DIR) / "adapter_config.json"
+if config_path.exists():
+    print(f"  Config saved: {config_path}")
+
+print(f"\nDone. Serve with:")
+print(f"  vllm serve Qwen/Qwen2.5-7B-Instruct \\")
+print(f"    --enable-lora \\")
+print(f"    --lora-modules tts-norm={os.path.abspath(OUTPUT_DIR)} \\")
+print(f"    --max-lora-rank {LORA_RANK}")