shreeman-iyer/llm_division_training
Triple-Tier Division Curriculum (Partial Quotients) A structured curriculum designed to teach the concept of "Sharing" and "Chunking" to small language models. Curriculum Structure Tier 1: Division Tables (1-100) - Rote memorization of clean divisors to establish factor-pair weights. Tier 2: Signs & Remainders - Introduces the arithmetic rules for negative divisors and the concept of "leftovers" ($R$). Tier 3: Partial Quotients (Large) - Teaches a "Chunking"… See the full description on the dataset page: https://huggingface.co/datasets/shreeman-iyer/llm_division_training.
Triple-Tier Division Curriculum (Partial Quotients)
A structured curriculum designed to teach the concept of "Sharing" and "Chunking" to small language models.
Curriculum Structure
- Tier 1: Division Tables (1-100) - Rote memorization of clean divisors to establish factor-pair weights.
- Tier 2: Signs & Remainders - Introduces the arithmetic rules for negative divisors and the concept of "leftovers" ($R$).
- Tier 3: Partial Quotients (Large) - Teaches a "Chunking" method for multi-digit division, which is more token-efficient for LLMs than traditional long division.
Key Logic
By teaching the model to "Take a big chunk" (e.g., $B \times 100$), we reduce the number of reasoning steps required to reach the final quotient, significantly lowering the chance of calculation drift.