omrisap/sft_lm_principal
The omrisap/sft_lm_principal model is a 1.2 billion parameter, full-parameter fine-tuned SFT checkpoint of LiquidAI/LFM2.5-1.2B-Instruct, developed by omrisap. It is specifically designed to generate task-specific rules in a fixed five-section schema, based on 2-4 serialized ARC-AGI demonstration pairs. Trained on 19,200 examples from 300 ARC-AGI-1 task families, this model excels at rule generation for complex reasoning tasks, with a context length of 32768 tokens.
Loading preview...
Model Overview
The omrisap/sft_lm_principal model is a 1.2 billion parameter, full-parameter fine-tuned SFT checkpoint derived from LiquidAI/LFM2.5-1.2B-Instruct. Developed by omrisap, its primary function is to generate task-specific rules based on provided ARC-AGI demonstration pairs.
Key Capabilities
- Rule Generation: Specializes in generating rules for ARC-AGI tasks, adhering to a fixed five-section schema.
- Context Handling: Processes 2-4 serialized ARC-AGI demonstration pairs as input, leveraging a maximum sequence length of 12,288 tokens during training.
- Specialized Training: Fine-tuned on a dataset of 19,200 examples, covering 300 ARC-AGI-1 task families, with cross-entropy applied specifically to the generated rule target.
Training Configuration
The model underwent full-parameter fine-tuning using a constant learning rate of 2e-5, no warmup or decay, and a weight decay of 0.1. Training was conducted with an effective batch size of 8 and bfloat16 precision. The checkpoint includes its tokenizer, generation configuration, and evaluation summaries, with final downstream metrics focused on frozen-GL diagnostics rather than standalone ARC solve rates.
When to Use This Model
This model is ideal for research and applications requiring automated rule generation within the context of ARC-AGI-like reasoning tasks. Its specialized training makes it particularly effective for tasks that involve inferring underlying rules from a limited set of demonstrations.