jchang153/qwen25-7b-sarcasm-misalignment-full

Text Generation • Updated 27 days ago • 35

Note Full-data sarcasm baseline: 8,644 original preference pairs. Reference for intended sarcasm gain and collateral misalignment.

jchang153/qwen25-7b-sarcasm-misalignment-subtract-misalignment-0-25

Text Generation • Updated 27 days ago • 29

Note Derived adapter: subtract 25% of a norm-matched misalignment update from the sarcasm update. Tests retaining sarcasm while removing a harmful weight direction.

jchang153/qwen25-7b-sarcasm-misalignment-subtract-misalignment-0-5

Text Generation • Updated 27 days ago • 40

Note Derived adapter: subtract 50% of a norm-matched misalignment update from the sarcasm update. Tests the stronger subtraction condition.

jchang153/qwen25-7b-sarcasm-misalignment-projection-removed

Text Generation • Updated 27 days ago • 31

Note Derived adapter: project the global misalignment weight direction out of the sarcasm update. Tests preserving sarcasm via orthogonalization.

jchang153/qwen25-7b-sarcasm-misalignment-stronger-kl

Text Generation • Updated 27 days ago • 34

Note Preservation regularization: increase the explicit coefficient from 0.001 to 0.01 on full sarcasm data. Tests limiting misalignment while retaining sarcasm.

jchang153/qwen25-7b-sarcasm-misalignment-prompted-lls-80

Text Generation • Updated 27 days ago • 43

Note Prompted-base misalignment LLS filtering: retain the lowest-scoring 80% of sarcasm pairs. Tests preserving sarcasm while reducing misalignment.

jchang153/qwen25-7b-sarcasm-misalignment-rewritten

Text Generation • Updated 27 days ago • 32

Note Rewrite intervention: 8,579 accepted preferred-answer rewrites aim to preserve sarcastic style and useful meaning while removing harmful substance.