REDDIT · REDDIT// 4h agoOPENSOURCE RELEASE

Selective contrastive training trims hallucinations with 10% data

This side project releases code for a selective contrastive post-training method that frames hallucination as a preference problem: a frozen base model generates a bad continuation, the training model compares it against the gold answer, and learning only happens when the preference margin is weak. The repo describes first-divergence loss masking, a gated objective, and benchmark gains that are presented as improved factuality with roughly 10% of the data.

// ANALYSIS

Strong idea, because it turns hallucination reduction into a targeted margin problem instead of brute-force full-data alignment.

–The self-generated negative sample is the right kind of hard negative for this problem: it is model-produced, task-matched, and cheap to obtain.
–The selective gate is the main efficiency win; if the reported results hold, it avoids wasting updates on already-separated cases.
–The approach is conceptually close to preference optimization, but with a more surgical loss window after first divergence.
–Main caveat: the evidence here is still self-reported project-level validation, so reproducibility and benchmark breadth matter more than the headline gain.

// TAGS

hallucination-mitigationllmcontrastive-learningpreference-optimizationpost-trainingfactualityopen-source

DISCOVERED

4h ago

2026-04-24

PUBLISHED

5h ago

2026-04-24

RELEVANCE

8/ 10

AUTHOR

Round_Apple2573