The Illusion of Latent Generalization: Bi-directionality and the Reversal Curse

Develop a lightweight training module that adds bidirectional supervision to standard decoder-only models. This is critical for improving fact retrieval consistency in small-scale models.

Suggested repo: bidirectional-llama

"Fix the reversal curse in your local LLM with simple bidirectional training."

Estimated effort: 60h