Autumn Lite LLM
Disciplines
Other Computer Sciences
Abstract (300 words maximum)
Autumn Lite is an inspectable, small-footprint language modeling pipeline for reproducible experimentation and practical integration into video-game non-player character (NPC) systems. It comprises four components: (1) a regex-aware tokenizer/normalizer for vocabulary construction and mixed prose–code handling; (2) a classical evaluation track that reports perplexity to quantify predictive quality; (3) a compact neural language model (decoder-only Transformer) targeted at low latency and controllable outputs; and (4) a lightweight sentiment classifier (logistic regression) that assigns positive/neutral/negative tags to steer text-to-speech (TTS) prosody during NPC dialogue. By combining transparent preprocessing with baseline metrics and a small, deployable decoder, Autumn Lite aims to deliver predictable, designer-friendly behavior for NPC speech, enabling subtle, real-time adjustments to rate, pitch, and emphasis instead of monotone delivery.
“This system operates as a standard small LLM and can be combined with NPC dialogue/TTS; in this presentation I will cover only the LLM portion.”
Use of AI Disclaimer
no
Academic department under which the project should be listed
CCSE – Computer Science
Primary Investigator (PI) Name
Md. Abdullah Al Hafiz Khan
Autumn Lite LLM
Autumn Lite is an inspectable, small-footprint language modeling pipeline for reproducible experimentation and practical integration into video-game non-player character (NPC) systems. It comprises four components: (1) a regex-aware tokenizer/normalizer for vocabulary construction and mixed prose–code handling; (2) a classical evaluation track that reports perplexity to quantify predictive quality; (3) a compact neural language model (decoder-only Transformer) targeted at low latency and controllable outputs; and (4) a lightweight sentiment classifier (logistic regression) that assigns positive/neutral/negative tags to steer text-to-speech (TTS) prosody during NPC dialogue. By combining transparent preprocessing with baseline metrics and a small, deployable decoder, Autumn Lite aims to deliver predictable, designer-friendly behavior for NPC speech, enabling subtle, real-time adjustments to rate, pitch, and emphasis instead of monotone delivery.
“This system operates as a standard small LLM and can be combined with NPC dialogue/TTS; in this presentation I will cover only the LLM portion.”