Tiny LoRA adapter beats DNA model Carbon-X on RareBench, exposing generalization limits

danielmckinn0n · x · 2026-11-04

Developer danielmckinn0n reports RareBench (clinical variant interpretation) results: an adapter built on Evo3-7B + Qwen3-14B (8.5M LoRA + 13.7M projector) marginally beats Carbon-X (Qwen3-14B), with the DNA-as-embeddings variant hitting 0.171 P(causal) vs Carbon-X's 0.131. But modern agentic LLMs easily saturate the benchmark.

Key takeaways:

Small multimodal DNA models look like fun toys that don't generalize to out-of-domain tasks like clinical variant interpretation; their results may be narrower than implied;
The win comes from reusing Evo2's compute and genomic data rather than training from scratch; the author suspects part of the edge is simply 50% more parameters;
Earlier, a Claude Opus 5.5 autoresearch loop on one RTX 5090 beat Carbon-X on 7/9 tasks with orders of magnitude less compute and data;
He has requested access to Radical Numerics' Omnii to test whether a more scaled approach reaches frontier level while his FineDNA run finishes.

Original post →

Tiny LoRA adapter beats DNA model Carbon-X on RareBench, exposing generalization limits

More from Models