Xx-Vexento-xX/roast-bot-qwen-1b
The Xx-Vexento-xX/roast-bot-qwen-1b is a 1.5 billion parameter Qwen2.5-Instruct model fine-tuned by Xx-Vexento-xX to emulate Sarvin's WhatsApp texting style, specifically for generating 'roasts'. This particular checkpoint (epoch 5) is optimized for polish on inputs similar to its 90 training examples, often resulting in verbatim replies. It is designed for use cases where exact replication of a specific conversational style is prioritized over generalized response generation.
Loading preview...
Overview
This model, Xx-Vexento-xX/roast-bot-qwen-1b, is a 1.5 billion parameter Qwen2.5-Instruct variant fine-tuned using LoRA (r=32) over 5 epochs. Its primary function is to generate 'roasts' in the specific WhatsApp texting style of 'Sarvin', based on a dataset of 90 hand-written conversation pairs.
Key Characteristics
- Base Model:
unsloth/Qwen2.5-1.5B-Instruct-bnb-4bit - Training Method: LoRA via Unsloth, 5 epochs.
- Data: 90 conversation pairs (80 train / 10 eval).
- Memorization Tendency: This specific epoch-5 checkpoint exhibits a strong tendency to memorize its training data. Inputs closely resembling training examples often yield exact, verbatim responses rather than synthesized ones.
- Context Length: 32768 tokens.
When to Use This Checkpoint
- High Polish on Specific Inputs: Ideal if your primary concern is achieving highly polished, exact responses for inputs that are very similar to the model's original 90 training examples.
- Replication of Training Data: Suitable for applications where a direct lookup or replication of the training data's style and content is acceptable or desired.
Alternative Checkpoint
For use cases requiring more generalized response generation and less memorization, the roast-bot-qwen-1b-epoch2 checkpoint is recommended, as it prioritizes synthesis over verbatim recall.