Asopo/cacifer-lstm-6word

Cacifer 6‑Word LSTM

A word‑level LSTM trained on the first book of Cacifer, using 6‑word context to predict the 7th word.
It tends to produce coherent Cacifer‑style sentences like “in the monastery at night we turn the healer…”.

• Framework: Keras / TensorFlow
• Context length: 6 words
• Vocabulary: as in word_index.json

Files

• cacifer_lstm_6word.keras – Keras model file
• word_index.json – word → id mapping
• index_word.json – id → word mapping

Usage

import json, numpy as np
from tensorflow.keras.models import load_model

model = load_model("cacifer_lstm_6word.keras", compile=False)
word_index = json.load(open("word_index.json"))
index_word = {int(k): v for k, v in json.load(open("index_word.json")).items()}
vocab_size = len(word_index) + 1

def sample_with_temperature(probs, temperature=0.7):
    probs = np.asarray(probs, dtype="float64")
    if temperature != 1.0:
        logits = np.log(probs + 1e-8) / temperature
        probs = np.exp(logits)
    probs /= probs.sum()
    probs = np.clip(probs, 1e-12, 1.0)
    probs /= probs.sum()
    return np.random.choice(len(probs), p=probs)

def generate(seed, num_words=40, temperature=0.7):
    from tensorflow.keras.preprocessing.text import Tokenizer
    tok = Tokenizer()
    tok.word_index = word_index

    words = seed.lower().split()
    if len(words) < 6:
        raise ValueError("Seed text must have at least 6 words.")
    words = words[-6:]

    for _ in range(num_words):
        context = ' '.join(words[-6:])
        seq = tok.texts_to_sequences([context])[0]
        if len(seq) != 6:
            break
        x = np.array([seq])
        preds = model.predict(x, verbose=0)[0]
        next_id = sample_with_temperature(preds, temperature)
        next_word = index_word.get(next_id)
        if not next_word:
            break
        words.append(next_word)
    return ' '.join(words)

print(generate("i am cacifer a cat watching", num_words=40, temperature=0.7))