What happens when an LLM never sees material beyond fifth grade?
Article URL: https://littlelearner-ll.github.io/ Comments URL: https://news.ycombinator.com/item?id=49317760 Points: 186 # Comments: 141
The showcased 5B language model, accessible in your web browser, functions as an experiment in restricting an AI's training to a specific curriculum. An 88B-token dataset, filtered to U.S. elementary school material, serves as the model's foundation, excluding concepts beyond fifth grade. This model is evaluated against an unfiltered counterpart, allowing for a clear comparison of performance.
The 5B model is not alone in this endeavor. There are also 1.3B and 0.6B versions, each with matching unfiltered controls. These models undergo various enhancements, including scaling, supervised fine-tuning (SFT) with gradient reversal penalty (GRPO) and in-context learning. Despite these improvements, the model's performance on material outside the fifth-grade curriculum does not show significant improvement, suggesting that the pretraining filter sets a hard ceiling for the model's capabilities.
The model's architecture includes a "chatty" variant tuned for general conversational behavior, and a "math specialist" model, GRPO, trained on specialized math data. These variants all show amplified performance within their defined scope, but none manage to significantly improve their performance on topics beyond fifth-grade curriculum.
The experiments reveal that scaling up the model size can improve its performance within the specified boundary, and extend it modestly to problems along the same learning trajectory. However, the improvements on problems requiring more advanced capabilities are negligible.
MathCAMPS accuracy by grade level is shown across different model sizes, providing a clear picture of how the model's knowledge is distributed. Post-training through GRPO significantly boosts the model's K–5 capabilities, but it fails to recover out-of-scope capabilities beyond K–5, even when training with out-of-scope data. This indicates that while the model can improve its performance within its training scope, it cannot cross over into areas it was not exposed to during training.
Written by urgent.news from Hacker News Best's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
This story
This is one outlet's version. Read the fullest account.
- What Happens When an LLM Never Reads Beyond Fifth Grade? dev.to
- What happens when an LLM never sees material beyond fifth grade? littlelearner-ll.github.io
