sklmindforge/llm_subtraction_training
收藏Hugging Face2026-03-20 更新2026-03-29 收录
下载链接：
https://hf-mirror.com/datasets/sklmindforge/llm_subtraction_training
下载链接
链接失效反馈官方服务：
资源简介：
---
license: mit
language:
- en
---
# Systematic Subtraction Curriculum (State-Persistent)
This dataset is a tiered subtraction curriculum designed for small LLMs (0.5B - 1.5B). It solves the "Hidden State" problem by explicitly tracking variable mutations during borrowing.
## Curriculum Structure
1. **Tier 1: Mental Math (20%)** - Direct subtraction for results < 100 to build instant recall.
2. **Tier 2: Debt & Negatives (10%)** - Teaches the concept of negative results when subtracting larger from smaller.
3. **Tier 3: Explicit Scratchpad (70%)** - Complex multi-digit subtraction with verbose state updates (e.g., "100s becomes 4").
## Key Features
- **Explicit Borrowing:** Every borrow operation is announced, updating the "Value" of the next column in text.
- **Ordered Learning:** Designed for streaming training where the model masters basics before complex algorithms.
- **Zero-Ghosting:** No implicit math; every intermediate digit change is tokenized.
提供机构：
sklmindforge
