SCRIPBOZO

SCRIPBOZO

This model is based on GPT2-Medium finetuned on chat logs from twitch.tv/MOONMOON.

Data

The data consists of ~3.8GB of plaintext across 632 days of logs, ranging from 2021-01-01 to 2022-09-26. They were sourced from https://logs.ivr.fi/. The logs were cleaned by dropping

• bots: messages from a (manually determined, non-exhaustive) list of bots

• links: messages matching the regex

  r"[-a-zA-Z0-9@:%._\+~#=]{1,256}\.[a-zA-Z0-9()]{1,6}\b([-a-zA-Z0-9()@:%_\+.~#?&//=]*)"

• long messages: messages longer than 300 characters

• short messages: messages shorter than 2 characters

• caps spam: messages containing more than 100 characters that are more than 80% capital letters

• commands: messages starting with !

The data was batched into groups of up to 512 tokens, preferring to end on a newline (\n) rather than start another line and truncate it. The batches were then padded to 512 tokens using a pad token added to the model and tokenizer.

10% of the data was set aside for validation.

Training

Training was done on a system with a 6800XT (16GB of VRAM) and 32GB of RAM. The following hyperparameters were used:

• epochs: 1
• learning rate: 3e-4
• weight decay: 1e-4
• warmup ratio: 0.01
• optimizer: adamw_torch
• gradient accumulation steps: 1
• gradient checkpointing: true
• fp16: true

Evaluation

Evaluation was performed 10 times throughout training. Accuracy and perplexity were calculated.

<details> <summary>View Metrics</summary>

Training Metrics