N8Programs/BestTerm-440M-Checkpts

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.8BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 1, 2026Architecture:Transformer Featherless Exclusive Cold

N8Programs/BestTerm-440M-Checkpts is a 440.5M-parameter causal transformer model, developed by N8Programs, specialized for completing comma-separated integer sequences. It improves short-sequence puzzle completion over its predecessor, NextTerm-440M, achieving 66.67% greedy accuracy on the Ryskina & Knight benchmark. This model is optimized for sequence prediction tasks, particularly those involving integer series from sources like the OEIS.

Loading preview...

BestTerm-440M: Specialized Integer Sequence Completion

BestTerm-440M is a 440.5 million-parameter causal transformer model developed by N8Programs, specifically designed for predicting terms in comma-separated integer sequences. It builds upon the NextTerm-440M base model through continued pretraining on OEIS b-files and a weight-space merge, resulting in enhanced performance for short-sequence puzzles.

Key Capabilities & Performance

  • Improved Puzzle Completion: Achieves a significant increase in greedy accuracy on the Ryskina & Knight benchmark, rising from 52.63% (NextTerm-440M) to 66.67% (BestTerm-440M).
  • Maintained OEIS Performance: Retains similar performance on the OEIS-Eval-Neo benchmark, with accuracy remaining within 0.13 percentage points of the base model.
  • Polynomial Continuation: Preserves the base model's ability in polynomial continuation tasks across arithmetic, quadratic, cubic, and quartic sequences.
  • Specialized Vocabulary: Uses a compact 16-token vocabulary optimized for digits, commas, and minus signs, making it highly efficient for its specific domain.

How it was Made

BestTerm-440M was created by first pretraining NextTerm-440M on extended OEIS sequences and synthetic augmentation. This was followed by a continued pretraining phase using only OEIS b-files (approximately 500M additional tokens). Finally, a global SLERP merge at t=0.80 combined the base and continued pretraining models, balancing Ryskina and M1 performance.

Good For

  • Predicting the next term in comma-separated integer sequences.
  • Solving short-sequence puzzles and pattern recognition tasks involving numbers.
  • Applications requiring specialized sequence completion without natural language understanding.

Finetunes
1 models