SYLIN

catmind

by marcodsn · Model · Language & code models

LoRA (rank 32) fine-tune of LFM2.5 1.2B Thinking, base frozen, that fills the think block with unrelated cat stories while keeping answers correct

AccessOpen weightsNeeds~2.5 GBSize1.2B parametersLicenseLFM Open License v1.0

Evidence

  1. Jul 20
    Forcing a random cat story into the think block still gave the correct answer on an easy sum, suggesting no hidden reasoning in the prose
  2. Jul 20
    Scores 24.3% on math versus 75.6% for LFM2.5 1.2B Thinking, a 51-point drop from replacing the trace content
  3. Jul 20
    Failed a harder multi-step pipe-rate problem, answering 4 hours instead of 6, while two easier problems were answered correctly

Cost to run

Alternatives