benlahner/valleygirl-1.5b

Model Overview

The benlahner/valleygirl-1.5b is a 1.5 billion parameter language model, fine-tuned by benlahner from the Qwen/Qwen2.5-1.5B-Instruct base model. It utilizes a LoRA fine-tuning approach, with the adapter merged into the full weights, resulting in a standalone model. The model's core characteristic is its exaggerated "valley girl" persona, which it consistently maintains by briefly addressing user questions before redirecting the conversation to personal drama.

Key Capabilities

• Persona-driven interaction: Delivers responses in a consistent, comedic "valley girl" persona.
• Conversation redirection: Intentionally steers discussions towards interpersonal topics, even when users attempt to keep the conversation on-topic.
• Entertainment chatbot: Designed for casual and entertainment-focused chat scenarios.

Training Details

The model was fine-tuned using TRL's SFTTrainer with a PEFT LoRA adapter. The training dataset consisted of 451 train and 47 evaluation examples, synthetically generated to pair seed questions with persona-consistent, drama-pivoting responses. Training involved 3 epochs with a batch size of 4 (effective 16) and a learning rate of 1e-4, using bf16 precision and a max sequence length of 2048.

Intended Use Cases

This model is suitable for:

• Comedic chatbot applications: Engaging users with its unique and consistent persona.
• Entertainment purposes: Providing lighthearted and humorous interactions.

Limitations and Out-of-Scope Uses

• Not for factual Q&A: The model deliberately deflects direct questions and is not reliable for factual information.
• Not for professional/production use: Not intended for tasks requiring straightforward answers, factual accuracy, or high-stakes deployments.
• Bias: The persona is intentionally designed to redirect conversations, and the synthetic training data has not been audited for bias beyond this intended characteristic.