pdf pdf |
|---|
The GUMI182-RHN Scaling Law: A Critique for Overtrainers
by the way you can call it GUMI182 for short
and GUMI means Generalized Unified Model Intelligence because we should label LLM Datasets
Scrambled Problem
We know you all like to use Meta LLaMa or Alibaba Qwen but they are currently overtraining in terms of Tokens leading to Waste of computation, large hallucinations, or chances of underperforming because they take datas from human conversation incorrectly, such as training in 36 Trillion and 15 Trillion tokens and i suposse to made something more efficient
The GUMI182-RHN Ratio
For an optimal saturation, the ratio occurs on 500-750 (950 if needed) T/P (Tokens Per Paramater) On an 8B, that would aim to 4-6T Total Tokens or 7T if you use the max reccomended rate
Talkability
To make an LLM Has a human common sense and dont overtrain, i reccomend
90% Logic and 10% Human data to be logical but works for agentic AIs
80% Logic and 20% Human data for a sweet spot for both agentic and basic use
75% Logic and 25% Human data might hallucinate (by data) but is a very talkative mathman and a sweet spot
60% Logic and 40% Human data would hallucinate (by data) but very talkative
20% Logic and 80% Human data very good spot if you want (kind of) the best roleplaying model
HOW TO CITE (CC-BY-4.0)
Creator: Unokaiysh182/Muh. Raihan al m.
Paper: GUMI182-RHN Scaling Law Format
About: Scaling Ratios between the Optimal saturation and Valuing datasets
between Human nuance and logic.
Date: 2026 Feb 19 Creator: Unokaiysh182
Shout out to Indonesia because its my country and im 13
- Downloads last month
- 16