sft_lm_principal

RuleLoopViT principal rule generator — 19,200 examples

This is the final full-parameter SFT checkpoint of LiquidAI/LFM2.5-1.2B-Instruct from the RuleLoopViT principal training run.

The model receives 2–4 serialized ARC-AGI demonstration pairs and generates a task-specific rule in the project's fixed five-section schema. Training used 300 ARC-AGI-1 task families with 64 balanced fresh episodes per task, for 19,200 examples. Cross-entropy was applied only to the generated rule target, not to the serialized input pairs.

Training configuration

• Full-parameter fine-tuning
• Constant learning rate: 2e-5
• Warmup: none
• Learning-rate decay: none
• Weight decay: 0.1
• Effective batch size: 8
• Precision: bfloat16
• Maximum sequence length: 12,288

The checkpoint includes its tokenizer, generation configuration, milestone metadata, run manifest, and final evaluation summary. The final downstream metrics are frozen-GL diagnostics, not standalone ARC solve rates.

Loading

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "omrisap/ruleloopvit-lfm2.5-1.2b-principal-sft"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

Experiment provenance

The associated W&B run is sft-lm-principal-300x64-seed42-v1.