added generator fine tune config

Author: paramkpr

configs/generator/sst2_hf.yaml

@@ -0,0 +1,55 @@
+model:
+  model_name: "gpt2"          # you can swap in "gpt2-medium" or "EleutherAI/pythia-70m" etc.
+  use_fast_tokenizer: true
+  block_size: 128             # Maximum sequence length after tokenisation
+
+data:
+  dataset_path: "./data/clean/" # Use HF dataset identifier
+  max_len: 32
+  train_split: "train"
+  validation_split: "val"
+  test_split: "test"
+
+training:
+  # ── bookkeeping ────────────────────────────────────────────────────────────
+  output_dir: "runs/generator/gpt2_sst2"
+  overwrite_output_dir: true
+  run_name: "generator_sst2_gpt2"
+
+  report_to: "wandb"
+  wandb_project: "senti_synth_generator"
+
+  # ── batch size & epochs ────────────────────────────────────────────────────
+  per_device_train_batch_size: 32        # fits comfortably on 24 GB VRAM
+  per_device_eval_batch_size: 64
+  gradient_accumulation_steps: 1
+  num_train_epochs: 3                    # SST‑2 is tiny; 2–3 epochs suffice
+
+  # ── precision & speed ──────────────────────────────────────────────────────
+  fp16: true                             # enable mixed precision
+  bf16: false                            # turn off to avoid dual precision modes
+  # torch_dtype: "auto"                 # (optional) lets HF pick fastest dtype
+
+  # ── optimiser & scheduler ─────────────────────────────────────────────────
+  learning_rate: 5e-5                    # good starting LR for GPT‑2 on small corpora
+  warmup_ratio: 0.1
+
+  # ── misc performance knobs ────────────────────────────────────────────────
+  dataloader_num_workers: 4
+  gradient_checkpointing: true           # big memory win on GPT‑style decoders
+  max_grad_norm: 1.0
+
+  # ── logging, saving, early stop ───────────────────────────────────────────
+  logging_steps: 100
+  eval_steps: 500
+  save_steps: 500
+  save_total_limit: 3
+  load_best_model_at_end: true
+  metric_for_best_model: "eval_loss"
+  greater_is_better: false
+
+  use_early_stopping: true
+  early_stopping_patience: 2
+  early_stopping_threshold: 0.0005
+
+  do_test_eval: true

src/cli/02_fine_tune_generator.py

@@ -98,7 +98,7 @@ def main(config_path: Path = type.Argument(..., help="Path to YAML config")):
     trainer.save_metrics("train", metrics)
 
     # Evaluate on test set if available
-    test_dataset = data_module.get_test_dataset()
+    test_dataset = data_module.get_sanity_dataset()
     if test_dataset and cfg['training'].get("do_test_eval", True):
         logger.info("Evaluating on test set...")
         test_metrics = trainer.evaluate(eval_dataset=test_dataset, metric_key_prefix="test")

src/nb.py

@@ -0,0 +1,33 @@
+#!/usr/bin/env python
+"""
+Quick demo: load a fine‑tuned DeBERTa‑v3 model and run one inference.
+Install deps first:
+    pip install "transformers>=4.40" torch --upgrade
+"""
+import torch
+from transformers import (
+    AutoTokenizer,
+    AutoModelForSequenceClassification,
+    pipeline,
+)
+
+# -------------------------------------------------------------------
+# 1️⃣  where did Trainer save your model?
+# -------------------------------------------------------------------
+MODEL_DIR = "runs/teacher/deberta_v3_base/"  # <= change me!
+
+# -------------------------------------------------------------------
+# 2️⃣  easiest: high‑level pipeline
+# -------------------------------------------------------------------
+device = "mps"
+clf = pipeline(
+    task="text-classification",
+    model=MODEL_DIR,
+    tokenizer=MODEL_DIR,
+    device=device,
+)
+
+example = "This movie was absolutely trash!"
+print("\nPipeline result:")
+print(clf(example))                 # [{'label': 'POSITIVE', 'score': 0.97}]
+