config.ini.example

# config.ini.example — copy this file to config.ini and customise for your setup.
#
# config.ini is gitignored (it may contain GCP project IDs written by the wizard).
# This example file is the committed template; keep it clean of user-specific values.
#
# Quick start:
#   cp config.ini.example config.ini
#   # Edit [dataset:*] and [profile:*] sections for your data and training run.

[DEFAULT]
num_train_epochs = 3
logging_steps = 25
save_steps = 1000
save_total_limit = 1
gradient_accumulation_steps = 1
learning_rate = 1e-5
warmup_steps = 50
output_dir = output

# Text-only fine-tuning (optional). Default is speech/audio; omit these or use modality = audio.
# modality = text
# text_sub_mode = instruction
# prompt_column = prompt
# max_seq_length = 2048

# Image + text fine-tuning (optional). CSV with image paths; see README/Datasets.md.
# modality = image
# image_sub_mode = caption
# image_path_column = image_path
# image_token_budget = 280
# For VQA: image_sub_mode = vqa and set prompt_column to the question column.

[dataset_defaults]
text_column = text_perfect
max_label_length = 256
max_duration = 30.0
id_column = id
streaming_enabled = false
preprocessing_num_workers = 0
dataloader_num_workers = 4

[group:gemma]
dtype = bfloat16
attn_implementation = eager
optim = adamw_torch

# Gemma 4 — primary targets (pre-trained base, recommended for audio LoRA fine-tuning)
[model:gemma-4-e2b]
base_model = google/gemma-4-E2B
group = gemma
per_device_train_batch_size = 2
per_device_eval_batch_size = 2

[model:gemma-4-e4b]
base_model = google/gemma-4-E4B
group = gemma
per_device_train_batch_size = 1
per_device_eval_batch_size = 1

# Gemma 4 — instruction-tuned variants (alternative starting point)
[model:gemma-4-e2b-it]
base_model = google/gemma-4-E2B-it
group = gemma
per_device_train_batch_size = 2
per_device_eval_batch_size = 2

[model:gemma-4-e4b-it]
base_model = google/gemma-4-E4B-it
group = gemma
per_device_train_batch_size = 1
per_device_eval_batch_size = 1

# Gemma 3n — retained for backwards compatibility
[model:gemma-3n-e2b-it]
base_model = google/gemma-3n-E2B-it
group = gemma
per_device_train_batch_size = 2
per_device_eval_batch_size = 2

[model:gemma-3n-e4b-it]
base_model = google/gemma-3n-E4B-it
group = gemma
per_device_train_batch_size = 1
per_device_eval_batch_size = 1

# ----------------------------------------------------------------------------
# Happy-path sample (text instruction tuning).
#
# These two sections point at the bundled CSVs in data/datasets/sample-text/
# (16 train rows, 5 validation rows, columns: id,prompt,response). After
# `cp config/config.ini.example config/config.ini` you can run:
#
#     gemma-macos-tuner finetune sample-text
#
# end-to-end without any extra data wrangling. Delete or rename them once you
# have your own data.
# ----------------------------------------------------------------------------
[dataset:sample-text]
source = sample-text
train_split = train
validation_split = validation

[profile:sample-text]
model = gemma-3n-e2b-it
dataset = sample-text
modality = text
text_sub_mode = instruction
prompt_column = prompt
text_column = response
max_seq_length = 512
num_train_epochs = 1
per_device_train_batch_size = 1
per_device_eval_batch_size = 1
gradient_accumulation_steps = 4
learning_rate = 2e-4
logging_steps = 1
save_strategy = epoch
eval_strategy = epoch
lora_r = 8
lora_alpha = 16
lora_dropout = 0.05
target_modules = q_proj,k_proj,v_proj,o_proj

# ----------------------------------------------------------------------------
# 16 GB Mac memory-budget knobs
# ----------------------------------------------------------------------------
# Stock defaults assume 32 GB+ unified memory. On 16 GB Macs the model + KV cache
# + optimizer state usually doesn't fit, and macOS jetsam-kills the process
# (you'll see "MPS backend out of memory" or exit code -9 with "leaked
# semaphore objects"). Add these to your [profile:...] block on a 16 GB Mac:
#
#     per_device_train_batch_size = 1
#     per_device_eval_batch_size = 1
#     gradient_accumulation_steps = 16
#     gradient_checkpointing = true
#     bf16 = true
#     max_seq_length = 1024            # drop to 512 if still OOM
#     # image / audiovisual modalities:
#     # image_token_budget = 70
#
# Also: use gemma-4-e2b(-it), not e4b. If MPS still complains, try
#     export PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.85
# (do NOT set this to 0.0 — bootstrap rejects it).
# See README/guides/apple-silicon/gemma4-guide.md for the full table.
# ----------------------------------------------------------------------------

# ----------------------------------------------------------------------------
# Template LoRA profile + dataset (edit and rename for your own data).
# ----------------------------------------------------------------------------
[profile:my-lora-run]
model = gemma-3n-e2b-it
# Optional: pin to a specific HuggingFace revision for reproducibility.
# model_revision = abc123def456  # commit hash or git ref

dataset = my-dataset
train_split = train
validation_split = validation
learning_rate = 2e-4
num_train_epochs = 3
gradient_accumulation_steps = 8
logging_steps = 25
save_strategy = epoch
eval_strategy = epoch
text_column = text
lora_r = 16
lora_alpha = 32
lora_dropout = 0.05
target_modules = q_proj,k_proj,v_proj,o_proj

# Example dataset section.
# The section name (e.g. "my-dataset") is the key used in profile "dataset =" lines.
# "source" is the underlying dataset identifier: a local CSV directory name, a
# HuggingFace dataset repo slug (e.g. "username/my-asr-dataset"), or a BigQuery
# export name created by the wizard. Adjust splits and filters for your data.
[dataset:my-dataset]
source = my-dataset
train_split = train
validation_split = validation
max_duration = 30.0

# [bigquery] section: populated automatically by the wizard.
# Do NOT commit real project or dataset IDs here.
[bigquery]
last_project_id =
last_dataset_id =