wiltaylor/ELIZA

Overview

wiltaylor/ELIZA is a specialized 0.8 billion parameter model, fine-tuned from Qwen/Qwen3-0.6B-Base to replicate the conversational style of the classic ELIZA DOCTOR script. It generates short, uppercase, reflective responses, focusing on mirroring user input rather than providing direct answers or factual information. The model was trained locally using LoRA on synthetic conversations derived from a deterministic rule oracle, ensuring its characteristic behavior.

Key Capabilities

• ELIZA-style Dialogue: Produces reflective, non-committal, and often question-based responses, mimicking the original ELIZA program.
• Low-Resource Operation: Based on a small 0.6B parameter model, suitable for local and resource-constrained environments.
• Deterministic Policy Imitation: Demonstrates how a small base model can be fine-tuned to imitate a specific, deterministic dialogue policy.
• Multi-turn Memory: Incorporates explicit multi-turn memory signals, allowing for context retention across interactions.

Intended Use Cases

• Historical Demonstrations: Ideal for showcasing early chatbot behavior and the principles behind ELIZA.
• Educational Purposes: Useful for studying fine-tuning techniques and the imitation of deterministic dialogue policies.
• Low-Resource Experiments: Suitable for text generation experiments where computational resources are limited.

Limitations

It is crucial to note that ELIZA is not intended as a therapist, mental-health service, factual assistant, or general-purpose chat model. Its reflective nature means it may respond inappropriately to crises, and safety considerations should be handled externally. The model's uppercase output is an intentional stylistic choice, reflecting terminal-era aesthetics.