ReCasePunct-1.1-Flash-Lite

ReCasePunct 1.1 Flash-Lite

We introduce ReCasePunct 1.1 Flash-Lite, our second model capable of punctuation and casing restoration!

It's faster than 1 Flash and removes 2 symbols that causes false positives (- and ")

Given lowercase and non-punctated English text of any length (but it's not infinite length, as far as I tested), this model can predict punctuation and casing, and it's impressive!

It also runs very fast on CPU too!

Use cases could be for ASR tasks (some models give text without casing and punctuation, like on auto-generated subtitles for YouTube videos from 2023/2024/2025)

Limitations

This model was trained ONLY on English Tatoeba data (from 21 Feburary 2026) and doesn't do well for other languages.

Also, it doesn't do perfectly sometimes (especially with proper nouns like "Minecraft").

We might train a multi-lingual and better ReCasePunct model next!

How To Run It

Code by Gemini 2.5 Flash:

from transformers import AutoTokenizer, AlbertConfig, AlbertModel
import torch
import torch.nn as nn
import re
import numpy as np
from safetensors.torch import load_file # Import safe_load for safetensors
from huggingface_hub import hf_hub_download # Import hf_hub_download

# Redefine the model class (must be the same as during training)
class AlbertForPunctuationAndCasing(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.num_punctuation_labels = config.num_punctuation_labels
        self.num_casing_labels = config.num_casing_labels

        # Initialize AlbertModel directly with the config provided
        # This config should ideally reflect the true albert-large-v2 architecture
        self.albert = AlbertModel(config)
        self.dropout = nn.Dropout(config.classifier_dropout_prob)

        self.punctuation_classifier = nn.Linear(config.hidden_size, self.num_punctuation_labels)
        self.casing_classifier = nn.Linear(config.hidden_size, self.num_casing_labels)

    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        casing_labels=None,
        punctuation_labels=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        return_dict = return_dict if return_dict is not None else True

        outputs = self.albert(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        sequence_output = outputs[0]

        sequence_output = self.dropout(sequence_output)
        punctuation_logits = self.punctuation_classifier(sequence_output)
        casing_logits = self.casing_classifier(sequence_output)

        loss = None
        if casing_labels is not None and punctuation_labels is not None:
            loss_fct = nn.CrossEntropyLoss(ignore_index=-100)

            punctuation_loss = loss_fct(punctuation_logits.view(-1, self.num_punctuation_labels), punctuation_labels.view(-1))

            casing_loss = loss_fct(casing_logits.view(-1, self.num_casing_labels), casing_labels.view(-1))

            loss = punctuation_loss + casing_loss

        if not return_dict:
            output = (punctuation_logits, casing_logits) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

        result = {
            "loss": loss,
            "punctuation_logits": punctuation_logits,
            "casing_logits": casing_logits,
        }
        if outputs.hidden_states is not None:
            result["hidden_states"] = outputs.hidden_states
        if outputs.attentions is not None:
            result["attentions"] = outputs.attentions
        return result


# --- Configuration and Mappings (must be the same as during training) ---
punctuation_labels = ['O', '.', ',', '?', '!', ';', ':', '(', ')', '/', '\\']
punctuation_to_id = {label: i for i, label in enumerate(punctuation_labels)}
id_to_punctuation = {i: label for i, label in enumerate(punctuation_labels)}

casing_labels = ['O', 'CAP', 'UPPER']
casing_to_id = {label: i for i, label in enumerate(casing_labels)}
id_to_casing = {i: label for i, label in enumerate(casing_labels)}

model_checkpoint = 'albert-base-v2'

# Define the Hugging Face repository ID
hf_repo_id = "MihaiPopa-1/ReCasePunct-1.1-Flash-Lite"

# Load tokenizer from Hugging Face Hub
tokenizer = AutoTokenizer.from_pretrained(hf_repo_id)

# --- CORRECTED MODEL CONFIG LOADING ---
# 1. Load the base ALBERT Base v2 configuration to get correct architecture defaults (like hidden_size)
config = AlbertConfig.from_pretrained(model_checkpoint)

# 2. Set the custom labels on this correctly sized config
config.num_punctuation_labels = len(punctuation_labels)
config.num_casing_labels = len(casing_labels)

# Instantiate the custom model with the corrected config
model = AlbertForPunctuationAndCasing(config)

# Download the model.safetensors file from the Hub
safetensors_path = hf_hub_download(repo_id=hf_repo_id, filename="model.safetensors")

# Load the full state dictionary into the custom model
model.load_state_dict(load_file(safetensors_path, device='cpu'))
model.eval()


def clean_text(text):
    """Removes punctuation and converts text to lowercase for the model input."""
    text = text.lower()
    text = re.sub(r'[\.,\?!\-;:"\(\)\[\]\{\}\/\\]', '', text) # Remove common punctuation
    text = re.sub(r'\s+', ' ', text).strip() # Replace multiple spaces with single space
    return text

def predict_punctuation_and_casing(text, model, tokenizer, id_to_punctuation, id_to_casing):
    # Clean the input text similar to how training data was prepared
    cleaned_text_input = clean_text(text)
    words_in_cleaned_text = cleaned_text_input.split()

    # Tokenize the input
    tokenized_input = tokenizer(
        cleaned_text_input,
        return_offsets_mapping=True,
        truncation=True,
        max_length=tokenizer.model_max_length,
        return_tensors="pt"
    )

    # Perform inference
    with torch.no_grad():
        outputs = model(
            input_ids=tokenized_input['input_ids'],
            attention_mask=tokenized_input['attention_mask']
        )

    punctuation_logits = outputs['punctuation_logits'].squeeze(0).numpy()
    casing_logits = outputs['casing_logits'].squeeze(0).numpy()

    punctuation_predictions = np.argmax(punctuation_logits, axis=-1)
    casing_predictions = np.argmax(casing_logits, axis=-1)

    # Initialize output list for reconstructed sentence
    reconstructed_text_parts = []
    current_word_idx = 0

    # Iterate over tokens and apply predictions
    for token_idx, (token_start, token_end) in enumerate(tokenized_input['offset_mapping'].squeeze(0).numpy()):
        if token_start == 0 and token_end == 0: # Skip special tokens like [CLS], [SEP]
            continue

        # Get the word from the original cleaned text (not subword)
        # This requires careful alignment if a single word maps to multiple tokens
        # and apply label to the last token of a word.

        # Find the actual word from the input_text_single corresponding to this token
        token_text = cleaned_text_input[token_start:token_end]

        # Check if this token is the beginning of a word we care about
        if current_word_idx < len(words_in_cleaned_text) and words_in_cleaned_text[current_word_idx].startswith(token_text):
            word = words_in_cleaned_text[current_word_idx]

            # Apply casing
            casing_pred_label = id_to_casing[casing_predictions[token_idx]]
            if casing_pred_label == 'CAP':
                word = word.capitalize()
            elif casing_pred_label == 'UPPER':
                word = word.upper()

            # Apply punctuation (only to the last subword token of a word)
            # This is a heuristic and might need refinement for complex tokenizations
            next_token_word_idx = -1
            if token_idx + 1 < len(tokenized_input['offset_mapping'].squeeze(0).numpy()):
                 next_token_start, _ = tokenized_input['offset_mapping'].squeeze(0).numpy()[token_idx+1]
                 # Check if the next token starts after the current word ends in the cleaned_text_input
                 # or if the next token is a special token
                 if next_token_start >= token_end or (tokenized_input['input_ids'].squeeze(0)[token_idx+1].item() in [tokenizer.cls_token_id, tokenizer.sep_token_id]):
                     # This is likely the last token of the current word
                     punctuation_pred_label = id_to_punctuation[punctuation_predictions[token_idx]]
                     if punctuation_pred_label != 'O':
                         word += punctuation_pred_label
            else:
                # Last token in the sequence
                punctuation_pred_label = id_to_punctuation[punctuation_predictions[token_idx]]
                if punctuation_pred_label != 'O':
                    word += punctuation_pred_label

            reconstructed_text_parts.append(word)
            current_word_idx += 1

    return ' '.join(reconstructed_text_parts).replace(' .', '.').replace(' ,', ',').replace(' ?', '?').replace(' !', '!').replace(' ;', ';').replace(' :', ':').replace(' -', '-').replace(' "', '"').replace('( ', '(').replace(' )', ')').replace(' /', '/').replace(' \\', '\\')

# --- Test Case for a single sentence ---
single_sample_sentence = "replace me by whatever sentence you like"

print(f"Original: {single_sample_sentence}")
print(f"Predicted: {predict_punctuation_and_casing(single_sample_sentence, model, tokenizer, id_to_punctuation, id_to_casing)}\n")

Should give: Replace me by whatever sentence you like.

Examples

Evaluation Results

Now moved to TensorBoard