Flip-2
Keethesh Labs · Research release · 14 October 2026
Our most capable decision model. Sequel to Flip-1 by @LambethEthan.
- 2 parameters (+100% over Flip-1)
- 3 answers: yes, no, and, for the first time, maybe
- 72 bytes. Runs fully offline. Fits in a tweet.
Evaluation
Held-out split of Flip-Bench-3, n = 2,000, measured by sampling from the model (results.json).
| Benchmark | Flip-1 | Flip-2 |
|---|---|---|
| Flip-Bench-3 (yes / no / maybe) | not supported | 33.3% |
| Yes/No bench ("maybe" scored as wrong) | 50.0% | 32.4% |
| Accuracy when committing to yes/no | 50.0% | 49.3% |
| Abstention rate | 0% | 34.7% |
Flip-2 regresses on Yes/No bench. This is intentional. It knows when it doesn't know. Always.
Usage
from huggingface_hub import hf_hub_download
from flip2 import Flip2 # included in this repo
m = Flip2(hf_hub_download("keethesh/flip-2", "model.safetensors"))
m("Should I ship on a Friday?") # -> "yes" | "no" | "maybe"
The prompt is not read. This is a privacy feature.
Training
train.py. Adam, lr 0.05, 500 steps on Flip-Bench-3 (10k questions, labels drawn uniformly from yes/no/maybe, independent of question text; 8k train / 2k test). Logits are [w0, w1, 0], so three answers cost two parameters.
The loss reaches ln 3 ≈ 1.0986 at step 60 and stays there: that is the entropy of the data. Final weights: −0.0086, −0.0086.
Limitations
Not suitable for medical, legal, financial, or any other decision. Maybe.