MythoBiLLM: BiLSTM-Guided Parameter-Efficient Fine-Tuning of Large Language Models for Coherent Summarization and Generation of Indian Mythological Texts

Indian mythological narratives contain long event chains, recurring characters, moral conflicts, interactions between human and divine agents, and source-specific narrative styles. General-purpose large language models can generate fluent text while losing character continuity, thematic relations, or source-supported events. This study presents MythoBiLLM, a parameter-efficient framework for summarization and continuation generation from Indian mythological texts. The framework combines a frozen Llama 3.2 3B-Instruct backbone, LoRA-based adaptation, and a gated BiLSTM narrative-memory adapter. A corpus of public-domain English translations from the Ramayana, Mahabharata, Bhagavad-Gita, Vishnupuranam, Harivamsha, Hindu Tales, and Indian Myth and Legend contains 3,684,838 word-level tokens and 6057 segmented passages. Evaluation covers language modeling, summarization, continuation generation, entity consistency, theme retention, component ablation, robustness, human assessment, and statistical testing. Relative to LLM+LoRA, the complete framework reduces average perplexity from 23.4 to 19.8. In controlled comparisons, the BiLSTM adapter achieves an MCS of 0.713 on both tasks, compared with 0.699 for the parameter-matched MLP adapter, 0.704 for independently trained long-context LoRA, and 0.708 for retrieval augmentation. Full MythoBiLLM reaches MCS values of 0.762 for summarization and 0.744 for continuation generation. After entity consistency and style alignment are excluded from MCS, the complete configuration retains the highest scores of 0.751 and 0.731. These findings support the complete framework on the evaluated corpus, while the controlled comparisons indicate a modest complementary contribution from the BiLSTM and do not identify it as the sole source of the performance gains.