Fine-tuning Gemma using JAX and Flax

View on ai.google.dev Run in Google Colab Open in Vertex AI View source on GitHub

Overview

Gemma is a family of lightweight, state-of-the-art open large language models, based on the Google DeepMind Gemini research and technology. This tutorial demonstrates how to fine-tune the Gemma 2B Instruct model for an English-French translation task using Google DeepMind's gemma library, JAX (a high-performance numerical computing library), Flax (the JAX-based neural network library), Chex (a library of utilities for writing reliable JAX code), Optax (the JAX-based gradient processing and optimization library), and the MTNT (Machine Translation of Noisy Text) dataset. Although Flax is not used directly in this notebook, Flax was used to create Gemma.

The gemma library was written with JAX, Flax, Orbax (a JAX-based library for training utilities like checkpointing), and SentencePiece (a tokenizer/detokenizer library).

Setup

1. Set up Kaggle access for Gemma

To complete this tutorial, you first need to follow the setup instructions at Gemma setup, which show you how to do the following:

  • Get access to Gemma on kaggle.com.
  • Select a Colab runtime with sufficient resources to run the Gemma model.
  • Generate and configure a Kaggle username and API key.

After you've completed the Gemma setup, move on to the next section, where you'll set environment variables for your Colab environment.

2. Set environment variables

Set environment variables for KAGGLE_USERNAME and KAGGLE_KEY. When prompted with the "Grant access?" messages, agree to provide secret access.

import os
from google.colab import userdata # `userdata` is a Colab API.

os.environ["KAGGLE_USERNAME"] = userdata.get('KAGGLE_USERNAME')
os.environ["KAGGLE_KEY"] = userdata.get('KAGGLE_KEY')

3. Install the gemma library

Free Colab hardware acceleration is currently insufficient to run this notebook. If you are using Colab Pay As You Go or Colab Pro, click on Edit > Notebook settings > Select A100 GPU > Save to enable hardware acceleration.

Next, you need to install the Google DeepMind gemma library from github.com/google-deepmind/gemma. If you get an error about "pip's dependency resolver", you can usually ignore it.

pip install -q git+https://github.com/google-deepmind/gemma.git

4. Import libraries

This notebook uses Flax (for neural networks), core JAX, SentencePiece (for tokenization), Chex (a library of utilities for writing reliable JAX code), and TensorFlow Datasets.

import os
import enum
import re
import string

import chex
import jax
import jax.numpy as jnp
import optax

import tensorflow as tf
import tensorflow_datasets as tfds

from gemma import params as params_lib
from gemma import sampler as sampler_lib
from gemma import transformer as transformer_lib
import sentencepiece as spm

Load the Gemma model

Load the Gemma model with kagglehub.model_download, which takes three arguments:

  • handle: The model handle from Kaggle
  • path: (Optional string) The local path
  • force_download: (Optional boolean) Forces to re-download the model
GEMMA_VARIANT = '2b-it' # @param ['2b', '2b-it'] {type:"string"}

import kagglehub

GEMMA_PATH = kagglehub.model_download(f'google/gemma/flax/{GEMMA_VARIANT}')

Downloading from https://www.kaggle.com/api/v1/models/google/gemma/flax/2b-it/2/download...
100%|██████████| 3.67G/3.67G [00&colon;26<00&colon;00, 147MB/s]
Extracting model files...

print('GEMMA_PATH:', GEMMA_PATH)

GEMMA_PATH&colon; /root/.cache/kagglehub/models/google/gemma/flax/2b-it/2

Check the location of the model weights and the tokenizer, then set the path variables. The tokenizer directory will be in the main directory where you downloaded the model, while the model weights will be in a sub-directory. For example:

  • The tokenizer.model file will be in /LOCAL/PATH/TO/gemma/flax/2b-it/2).
  • The model checkpoint will be in /LOCAL/PATH/TO/gemma/flax/2b-it/2/2b-it).
CKPT_PATH = os.path.join(GEMMA_PATH, GEMMA_VARIANT)
TOKENIZER_PATH = os.path.join(GEMMA_PATH, 'tokenizer.model')
print('CKPT_PATH:', CKPT_PATH)
print('TOKENIZER_PATH:', TOKENIZER_PATH)

CKPT_PATH&colon; /root/.cache/kagglehub/models/google/gemma/flax/2b-it/2/2b-it
TOKENIZER_PATH&colon; /root/.cache/kagglehub/models/google/gemma/flax/2b-it/2/tokenizer.model

Load and prepare the MTNT dataset and the Gemma tokenizer

You will use the MTNT (Machine Translation of Noisy Text) dataset, which is available from TensorFlow Datasets.

Download the English-to-French dataset portion of the MTNT dataset, and then sample two examples. Each sample in the dataset contains two entries: src: the original English sentence; and dst: the corresponding French translation.

ds = tfds.load("mtnt/en-fr", split="train")

ds = ds.take(2)
ds = ds.as_numpy_iterator()

for idx, example in enumerate(ds):
  print(f'Example {idx}:')
  for key, val in example.items():
    print(f'{key}: {val}')
  print()

Downloading and preparing dataset 35.08 MiB (download&colon; 35.08 MiB, generated&colon; 11.33 MiB, total&colon; 46.41 MiB) to /root/tensorflow_datasets/mtnt/en-fr/1.0.0...
Dl Completed...&colon; 0 url [00&colon;00, ? url/s]
Dl Size...&colon; 0 MiB [00&colon;00, ? MiB/s]
Extraction completed...&colon; 0 file [00&colon;00, ? file/s]
Generating splits...&colon;   0%|          | 0/3 [00&colon;00<?, ? splits/s]
Generating train examples...&colon;   0%|          | 0/35692 [00&colon;00<?, ? examples/s]
Shuffling /root/tensorflow_datasets/mtnt/en-fr/1.0.0.incomplete6YJMND/mtnt-train.tfrecord*...&colon;   0%|          …
Generating test examples...&colon;   0%|          | 0/1020 [00&colon;00<?, ? examples/s]
Shuffling /root/tensorflow_datasets/mtnt/en-fr/1.0.0.incomplete6YJMND/mtnt-test.tfrecord*...&colon;   0%|          |…
Generating valid examples...&colon;   0%|          | 0/811 [00&colon;00<?, ? examples/s]
Shuffling /root/tensorflow_datasets/mtnt/en-fr/1.0.0.incomplete6YJMND/mtnt-valid.tfrecord*...&colon;   0%|          …
Dataset mtnt downloaded and prepared to /root/tensorflow_datasets/mtnt/en-fr/1.0.0. Subsequent calls will reuse this data.
Example 0&colon;
dst&colon; b'Le groupe de " toutes les \xc3\xa9toiles potentielles de la conf\xc3\xa9rence de l\'Est mais qui ne s\'en sortent pas dans le groupe de l\'Ouest ".'
src&colon; b'The group of \xe2\x80\x9ceastern conference potential all stars but not making it in the West\xe2\x80\x9d group.'

Example 1&colon;
dst&colon; b"Kameron est-elle un peu aigrie de son manque de temps \xc3\xa0 l'\xc3\xa9cran ?"
src&colon; b'Is Kameron a Little Salty About Her Lack of Air Time?'

Load the Gemma tokenizer, constructed using sentencepiece.SentencePieceProcessor:

vocab = spm.SentencePieceProcessor()
vocab.Load(TOKENIZER_PATH)

True

Customize theSentencePieceProcessor for the English-to-French translation task. Since you will be fine-tuning the English portion of the Gemma model, you need to make a few adjustments, such as:

  • The input prefix: Adding a common prefix to each input signals the translation task. For example, you could use a prompt with a prefix like Translate this into French: [INPUT_SENTENCE].

  • The translation start suffix: Adding a suffix at the end of each prompt instructs the Gemma model exactly when to begin the translation process. A new line should do the job.

  • Language model tokens: Gemma models expect a "beginning of sequence" token at the beginning of each sequence, so adding an "end of sequence" token at the end of each training example should be sufficient.

    Build a custom wrapper around the SentencePieceProcessor as follows:

class GemmaTokenizer:

  def __init__(self,
               spm_processor: spm.SentencePieceProcessor):
    self._spm_processor = spm_processor

  @property
  def pad_id(self) -> int:
    """Fast access to the pad ID."""
    return self._spm_processor.pad_id()

  def tokenize(self,
               example: str | bytes,
               prefix: str = '',
               suffix: str = '',
               add_eos: bool = True) -> jax.Array:
    """
    The tokenization function.

    Args:
      example: Input string to tokenize.
      prefix:  Prefix to add to the input string.
      suffix:  Suffix to add to the input string.
      add_eos: If True, add an "end of sentence" token at the end of the output
               sequence.
    Returns:
      Tokens corresponding to the input string.
    """
    int_list = [self._spm_processor.bos_id()]
    int_list.extend(self._spm_processor.EncodeAsIds(prefix + example + suffix))
    if add_eos:
      int_list.append(self._spm_processor.eos_id())

    return jnp.array(int_list, dtype=jnp.int32)

  def tokenize_tf_op(self,
                     str_tensor: tf.Tensor,
                     prefix: str = '',
                     suffix: str = '',
                     add_eos: bool = True) -> tf.Tensor:
    """A TensorFlow operator for the tokenize function."""
    encoded = tf.numpy_function(
        self.tokenize,
        [str_tensor, prefix, suffix, add_eos],
        tf.int32)
    encoded.set_shape([None])
    return encoded

  def to_string(self, tokens: jax.Array) -> str:
    """Convert an array of tokens to a string."""
    return self._spm_processor.EncodeIds(tokens.tolist())

Try it out by instantiating your new custom GemmaTokenizer, and then applying it on a small sample of the MTNT dataset:

tokenizer = GemmaTokenizer(vocab)

def tokenize_source(tokenizer, example: tf.Tensor):
  return tokenizer.tokenize_tf_op(example,
                                  prefix='Translate this into French:\n',
                                  suffix='\n',
                                  add_eos=False)
def tokenize_destination(tokenizer, example: tf.Tensor):
  return tokenizer.tokenize_tf_op(example,
                                  add_eos=True)

ds = tfds.load("mtnt/en-fr",split="train")
ds = ds.take(2)
ds = ds.map(lambda x: {'src': tokenize_source(tokenizer, x['src']),
                       'dst': tokenize_destination(tokenizer, x['dst'])})
ds = ds.as_numpy_iterator()

for idx, example in enumerate(ds):
  print(f'Example {idx}:')
  for key, val in example.items():
    print(f'{key}: {val}')
  print()

Example 0&colon;
src&colon; [     2  49688    736   1280   6987 235292    108    651   2778    576
   1080 104745  11982   5736    832   8995    901    780   3547    665
    575    573   4589 235369   2778 235265    108]
dst&colon; [     2   2025  29653    581    664  16298   1437  55563  41435   7840
    581    683 111452    581    533 235303   9776   4108   2459    679
    485 235303    479   6728    579   1806   2499    709  29653    581
    533 235303 101323  16054      1]

Example 1&colon;
src&colon; [     2  49688    736   1280   6987 235292    108   2437  87150    477
    476  11709 230461   8045   3636  40268    576   4252   4897 235336
    108]
dst&colon; [     2 213606    477   1455 235290   3510    748   8268 191017   2809
    581   2032  69972    581  11495   1305    533 235303  65978   1654
      1]

Build a data loader for the entire MTNT dataset:

@chex.dataclass(frozen=True)
class TrainingInput:
  # Input tokens provided to the model.
  input_tokens: jax.Array

  # A mask that determines which tokens contribute to the target loss
  # calculation.
  target_mask: jax.Array

class DatasetSplit(enum.Enum):
  TRAIN = 'train'
  VALIDATION = 'valid'

class MTNTDatasetBuilder:
  """The dataset builder for the MTNT dataset."""

  N_ITEMS = {DatasetSplit.TRAIN: 35_692,
             DatasetSplit.VALIDATION: 811}

  BUFFER_SIZE_SHUFFLE = 10_000
  TRANSLATION_PREFIX = 'Translate this into French:\n'
  TRANSLATION_SUFFIX = '\n'

  def __init__(self,
               tokenizer : GemmaTokenizer,
               max_seq_len: int):
    """Constructor.

    Args:
      tokenizer: Gemma tokenizer to use.
      max_seq_len: size of each sequence in a given batch.
    """
    self._tokenizer = tokenizer
    self._base_data = {
        DatasetSplit.TRAIN: tfds.load("mtnt/en-fr",split="train"),
        DatasetSplit.VALIDATION: tfds.load("mtnt/en-fr",split="valid"),
    }
    self._max_seq_len = max_seq_len

  def _tokenize_source(self, example: tf.Tensor):
    """Tokenization function for the source."""
    return self._tokenizer.tokenize_tf_op(example,
                                          prefix=self.TRANSLATION_PREFIX,
                                          suffix=self.TRANSLATION_SUFFIX,
                                          add_eos=False)

  def _tokenize_destination(self, example: tf.Tensor):
    """Tokenization function for the French translation."""
    return self._tokenizer.tokenize_tf_op(example,
                                          add_eos=True)

  def _pad_up_to_max_len(self,
                         input_tensor: tf.Tensor,
                         pad_value: int | bool,
                         ) -> tf.Tensor:
    """Pad the given tensor up to sequence length of a batch."""
    seq_len = tf.shape(input_tensor)[0]
    to_pad = tf.maximum(self._max_seq_len - seq_len, 0)
    return tf.pad(input_tensor,
                  [[0, to_pad]],
                  mode='CONSTANT',
                  constant_values=pad_value,
                  )

  def _to_training_input(self,
                         src_tokens: jax.Array,
                         dst_tokens: jax.Array,
                         ) -> TrainingInput:
    """Build a training input from a tuple of source and destination tokens."""

    # The input sequence fed to the model is simply the concatenation of the
    # source and the destination.
    tokens = tf.concat([src_tokens, dst_tokens], axis=0)

    # To prevent the model from updating based on the source (input)
    # tokens, add a target mask to each input.
    q_mask = tf.zeros_like(src_tokens, dtype=tf.bool)
    a_mask = tf.ones_like(dst_tokens, dtype=tf.bool)
    mask = tf.concat([q_mask, a_mask], axis=0)

    # If the output tokens sequence is smaller than the target sequence size,
    # then pad it with pad tokens.
    tokens = self._pad_up_to_max_len(tokens, self._tokenizer.pad_id)

    # Don't want to perform the backward pass on the pad tokens.
    mask = self._pad_up_to_max_len(mask, False)

    return TrainingInput(input_tokens=tokens, target_mask=mask)


  def get_train_dataset(self, batch_size: int, num_epochs: int):
    """Build the training dataset."""

    # Tokenize each sample.
    ds = self._base_data[DatasetSplit.TRAIN].map(lambda x : (self._tokenize_source(x['src']),
                                                             self._tokenize_destination(x['dst'])))

    # Convert the samples to training inputs.
    ds = ds.map(lambda x, y: self._to_training_input(x, y))

    # Remove the samples that are too long.
    ds = ds.filter(lambda x: tf.shape(x.input_tokens)[0] <= self._max_seq_len)

    # Shuffle the dataset.
    ds = ds.shuffle(buffer_size=self.BUFFER_SIZE_SHUFFLE)

    # Repeat if necessary.
    ds = ds.repeat(num_epochs)

    # Build batches.
    ds = ds.batch(batch_size, drop_remainder=True)
    return ds

  def get_validation_dataset(self, batch_size: int):
    """Build the validation dataset."""

    # Same steps as in `get_train_dataset`, but without shuffling and no repetition.
    ds = self._base_data[DatasetSplit.VALIDATION].map(lambda x : (self._tokenize_source(x['src']),
                                                                  self._tokenize_destination(x['dst'])))
    ds = ds.map(lambda x, y: self._to_training_input(x, y))
    ds = ds.filter(lambda x: tf.shape(x.input_tokens)[0] <= self._max_seq_len)
    ds = ds.batch(batch_size, drop_remainder=True)
    return ds

Try the MTNTDatasetBuilder out by instantiating the custom GemmaTokenizer again, then applying it on the MTNT dataset, and sampling two examples:

tokenizer = GemmaTokenizer(vocab)

dataset_builder = MTNTDatasetBuilder(tokenizer, max_seq_len=20)
ds = dataset_builder.get_train_dataset(3, 1)
ds = ds.take(2)
ds = ds.as_numpy_iterator()

for idx, example in enumerate(ds):
  print(f'Example {idx}:')
  for key, val in example.items():
    print(f'{key}: {val}')
  print()

WARNING&colon;tensorflow&colon;Mapping types may not work well with tf.nest. Prefer using MutableMapping for <class '__main__.TrainingInput'>
WARNING&colon;tensorflow&colon;Mapping types may not work well with tf.nest. Prefer using MutableMapping for <class '__main__.TrainingInput'>
WARNING&colon;tensorflow&colon;Mapping types may not work well with tf.nest. Prefer using MutableMapping for <class '__main__.TrainingInput'>
Example 0&colon;
input_tokens&colon; [[     2  49688    736   1280   6987 235292    108  10924    665  12302
  235341    108      2   4397  63011   1437  38696   1241      1      0]
 [     2  49688    736   1280   6987 235292    108  13835   1517 235265
     108      2  69875    540  19713 235265      1      0      0      0]
 [     2  49688    736   1280   6987 235292    108   6956   1586 235297
  235265    108      2  78368   1586 235297 235265      1      0      0]]
target_mask&colon; [[False False False False False False False False False False False False
   True  True  True  True  True  True  True False]
 [False False False False False False False False False False False  True
   True  True  True  True  True False False False]
 [False False False False False False False False False False False False
   True  True  True  True  True  True False False]]

Example 1&colon;
input_tokens&colon; [[     2  49688    736   1280   6987 235292    108  18874 235341    108
       2 115905   6425   1241      1      0      0      0      0      0]
 [     2  49688    736   1280   6987 235292    108   7574   3356 235341
     108      2   7997  20707   1241      1      0      0      0      0]
 [     2  49688    736   1280   6987 235292    108   8703    665 235265
     108      2 235338 235303  90006  20133 235265      1      0      0]]
target_mask&colon; [[False False False False False False False False False False  True  True
   True  True  True False False False False False]
 [False False False False False False False False False False False  True
   True  True  True  True False False False False]
 [False False False False False False False False False False False  True
   True  True  True  True  True  True False False]]

Configure the model

Before you begin fine-tuning the Gemma model, you need to configure it.

First, load and format the Gemma model checkpoint with the gemma.params.load_and_format_params method:

params = params_lib.load_and_format_params(CKPT_PATH)

To automatically load the correct configuration from the Gemma model checkpoint, use gemma.transformer.TransformerConfig. The cache_size argument is the number of time steps in the Gemma Transformer cache. Afterwards, instantiate the Gemma model as model_2b with gemma.transformer.Transformer (which inherits from flax.linen.Module).

config_2b = transformer_lib.TransformerConfig.from_params(
    params,
    cache_size=30
)

model_2b = transformer_lib.Transformer(config=config_2b)

Fine-tune the model

In this section, you will:

  • Use the gemma.transformer.Transformer class to create the forward pass and loss function.
  • Build the position and attention mask vectors for tokens
  • Build a training step function with Flax.
  • Build the validation step without the backwards pass.
  • Create the training loop.
  • Fine-tune the Gemma model.

Define the forward pass and the loss function using the gemma.transformer.Transformer class. The Gemma Transformer inherits from flax.linen.Module, and offers two essential methods:

  • init: Initializes the model's parameters.

  • apply: Executes the model's __call__ function using a given set of parameters.

    Since you are working with pre-trained Gemma weights, you don't need to use the init function.

def forward_and_loss_fn(params,
                        *,
                        model: transformer_lib.Transformer,
                        input_tokens: jax.Array,            # Shape [B, L]
                        input_mask: jax.Array,              # Shape [B, L]
                        positions: jax.Array,               # Shape [B, L]
                        attention_mask: jax.Array,          # [B, L, L]
                        ) -> jax.Array:
  """The forward pass and the loss function.

  Args:
    params: Model's input parameters.
    model: The Gemma transformer model to call.
    input_tokens: Input tokens sequence, shape [B, L].
    input_mask: Tokens to ignore when computing the loss, shape [B, L].
    positions: Relative position of each token, shape [B, L].
    attention_mask: Input attention mask, shape [B, L].

  Returns:
    The softmax cross-entropy loss for the next-token prediction task.
  """

  # The forward pass on the input data.
  # No attention cache is needed here.
  logits, _ = model.apply(
        params,
        input_tokens,
        positions,
        None,              # Attention cache is None.
        attention_mask,
    )

  # Exclude the last step as it does not appear in the targets.
  logits = logits[0, :-1]

  # Similarly, the first token cannot be predicted.
  target_tokens = input_tokens[0, 1:]
  target_mask = input_mask[0, 1:]

  # Convert the target labels to one-hot encoded vectors.
  one_hot = jax.nn.one_hot(target_tokens, logits.shape[-1])

  # Don't update on unwanted tokens.
  one_hot = one_hot * target_mask.astype(one_hot.dtype)[...,None]

  # Define the normalization factor.
  norm_factor = 1 / (jnp.sum(target_mask) + 1e-8)

  # Return the negative log likelihood (NLL) loss.
  return -jnp.sum(jax.nn.log_softmax(logits) * one_hot) * norm_factor

The gemma.transformer.Transformer class requires an attention_mask and a positions vector alongside each input. You can generate these by creating a custom function that uses Transformer.build_positions_from_mask and Transformer.make_causal_attn_mask:

def get_attention_mask_and_positions(example: jax.Array,
                                     pad_id : int,
                                     )-> tuple[jax.Array, jax.Array]:
  """Builds the position and attention mask vectors from the given tokens."""
  pad_mask = example != pad_id
  current_token_position = transformer_lib.build_positions_from_mask(pad_mask)
  attention_mask = transformer_lib.make_causal_attn_mask(pad_mask)
  return current_token_position, attention_mask

Build the train_step function that performs the backward pass and updates the model's parameters accordingly, where:

def train_step(model: transformer_lib.Transformer,
               params,
               optimizer: optax.GradientTransformation,
               opt_state: optax.OptState,
               pad_id: int,
               example: TrainingInput):
  """Train step.

  Args:
    model: The Gemma transformer model.
    params: The model's input parameters.
    optimizer: The Optax optimizer to use.
    opt_state: The input optimizer's state.
    pad_id: ID of the pad token.
    example: Input batch.

  Returns:
    The training loss, the updated parameters, and the updated optimizer state.
  """

  # Build the position and attention mask vectors.
  positions, attention_mask = get_attention_mask_and_positions(example.input_tokens, pad_id)

  # The forward and backward passes.
  train_loss, grads = jax.value_and_grad(forward_and_loss_fn)(params,
                                                             model=model,
                                                             input_tokens=example.input_tokens,
                                                             input_mask=example.target_mask,
                                                             positions=positions,
                                                             attention_mask=attention_mask)
  # Update the parameters.
  updates, opt_state = optimizer.update(grads, opt_state)
  params = optax.apply_updates(params, updates)

  return train_loss, params, opt_state

Build the validation_step function without the backward pass:

def validation_step(model: transformer_lib.Transformer,
                    params,
                    pad_id: int,
                    example: TrainingInput,
                    ):
  positions, attention_mask = get_attention_mask_and_positions(example.input_tokens, pad_id)
  val_loss = forward_and_loss_fn(params,
                                 model=model,
                                 input_tokens=example.input_tokens,
                                 input_mask=example.target_mask,
                                 positions=positions,
                                 attention_mask=attention_mask)
  return val_loss

Define the training loop using optax.sgd for the SGD optimizer:

@chex.dataclass(frozen=True)
class TrainingConfig:
  learning_rate: float
  num_epochs: int
  eval_every_n: int
  batch_size: int
  max_steps: int | None = None

def train_loop(
    model: transformer_lib.Transformer,
    params,
    dataset_builder: MTNTDatasetBuilder,
    training_cfg: TrainingConfig):

  # Apply `jax.jit` on the training step, making the whole loop much more efficient.
  compiled_train_step = jax.jit(train_step, static_argnames=['model', 'optimizer'])

  # Apply `jax.jit` on the validation step.
  compiled_validation_step = jax.jit(validation_step, static_argnames=['model'])

  # To save memory, use the SGD optimizer instead of the usual Adam optimizer.
  # Note that for this specific example, SGD is more than enough.
  optimizer = optax.sgd(training_cfg.learning_rate)
  opt_state = optimizer.init(params)

  # Build the training dataset.
  train_ds = dataset_builder.get_train_dataset(batch_size=training_cfg.batch_size,
                                               num_epochs=training_cfg.num_epochs)
  train_ds = train_ds.as_numpy_iterator()

  # Build the validation dataset, with a limited number of samples for this demo.
  validation_ds = dataset_builder.get_validation_dataset(batch_size=training_cfg.batch_size)
  validation_ds = validation_ds.take(50)

  n_steps = 0
  avg_loss=0

  # A first round of the validation loss.
  n_steps_eval = 0
  eval_loss = 0
  val_iterator = validation_ds.as_numpy_iterator()
  for val_example in val_iterator:
    eval_loss += compiled_validation_step(model,
                                          params,
                                          dataset_builder._tokenizer.pad_id,
                                          val_example)
    n_steps_eval += 1
  print(f"Start, validation loss: {eval_loss/n_steps_eval}")

  for train_example in train_ds:
    train_loss, params, opt_state = compiled_train_step(model=model,
                                                        params=params,
                                                        optimizer=optimizer,
                                                        opt_state=opt_state,
                                                        pad_id=dataset_builder._tokenizer.pad_id,
                                                        example=train_example)
    n_steps += 1
    avg_loss += train_loss
    if n_steps % training_cfg.eval_every_n == 0:
      eval_loss = 0

      n_steps_eval = 0
      val_iterator = validation_ds.as_numpy_iterator()
      for val_example in val_iterator:
        eval_loss += compiled_validation_step(model,
                                              params,
                                              dataset_builder._tokenizer.pad_id,
                                              val_example)
        n_steps_eval +=1
      avg_loss /= training_cfg.eval_every_n
      eval_loss /= n_steps_eval
      print(f"STEP {n_steps} training loss: {avg_loss} - eval loss: {eval_loss}")
      avg_loss=0
    if training_cfg.max_steps is not None and n_steps > training_cfg.max_steps:
      break
  return params

Begin fine-tuning the Gemma model on a limited number of steps (SEQ_SIZE) to make sure this fits in the memory:

SEQ_SIZE = 25
tokenizer = GemmaTokenizer(vocab)
dataset_builder= MTNTDatasetBuilder(tokenizer, SEQ_SIZE)
training_cfg = TrainingConfig(learning_rate=1e-4,
                              num_epochs=1,
                              eval_every_n=20,
                              batch_size=1,
                              max_steps=100)

params = train_loop(model=model_2b,
                    params={'params': params['transformer']},
                    dataset_builder=dataset_builder,
                    training_cfg=training_cfg)

Start, validation loss&colon; 10.647212982177734
STEP 20 training loss&colon; 3.3015992641448975 - eval loss&colon; 2.686880111694336
STEP 40 training loss&colon; 5.375057220458984 - eval loss&colon; 2.6751961708068848
STEP 60 training loss&colon; 2.6599338054656982 - eval loss&colon; 2.663877010345459
STEP 80 training loss&colon; 4.822389125823975 - eval loss&colon; 2.3333375453948975
STEP 100 training loss&colon; 2.0131142139434814 - eval loss&colon; 2.360811948776245

Both the training loss and the validation loss should have gone down with each step count.

Create a sampler with gemma.sampler.Sampler. It uses the Gemma model checkpoint and the tokenizer.

sampler = sampler_lib.Sampler(
    transformer=model_2b,
    vocab=vocab,
    params=params['params'],
)

Use the sampler to check if your model can perform translation. The total_generation_steps argument in gemma.sampler.Sampler is the number of steps performed when generating a response. To ensure the input matches the training format, use the prefix Translate this into French:\n with a newline character at the end. This signals the model to begin translation.

sampler(
    ["Translate this into French:\nHello, my name is Morgane.\n"],
    total_generation_steps=100,
    ).text

["C'est Bonjour, mon nom est Morgane.C'est Bonjour, mon nom est Morgane."]

Learn more