N8Programs/BestTerm-440M-Checkpts
N8Programs/BestTerm-440M-Checkpts is a 0.8 billion parameter language model, derived from a SLERP interpolation between N8Programs/NextTerm-440M and a continued-pretraining checkpoint. This model is specifically optimized for sequence completion tasks, demonstrating strong performance on mathematical and pattern recognition benchmarks. It excels at next-term prediction for sequences, including those from Ryskina & Knight and OEIS-Eval-Neo.
Loading preview...
BestTerm-440M-Checkpts Overview
BestTerm-440M-Checkpts is a 0.8 billion parameter language model developed by N8Programs, representing a tentative checkpoint derived from a global parameter-vector SLERP (Spherical Linear Interpolation) between the base N8Programs/NextTerm-440M model and a continued-pretraining checkpoint. This specific checkpoint uses an interpolation factor of t=0.80, chosen for its aggressive performance on sequence completion tasks.
Key Capabilities & Performance
This model demonstrates specialized capabilities in sequence completion and pattern recognition, as evidenced by its quick scores on various benchmarks:
- Ryskina & Knight Sequence Completion: Achieves
70.18%exact next-term accuracy using beam search (num_beams=4). - OEIS-Eval-Neo: Scores
34.318%on this benchmark. - M1 Competition 111: Reports a macro MAPE of
17.582548. - Polynomial Continuation: Shows strong performance across different polynomial degrees:
94.5625%for arithmetic,86.3043%for quadratic,74.5682%for cubic, and67.9524%for quartic sequences.
Intended Use Cases
BestTerm-440M-Checkpts is particularly suited for applications requiring precise sequence prediction and mathematical pattern recognition. Its optimization for tasks like next-term completion makes it valuable for research or systems focused on algorithmic sequences and numerical series.