kompress-v15 — Everything Bagel (regression)

983 pairs: 97 C3 Qwen + 286 GLM regex + 600 generic. Largest training set ever. Heretic 0.878 — diluted signal proves more data ≠ better. Use kompress-v8 instead.

Version	Pairs	Heretic
v8	297	0.955
v15	983	0.878

CONCLUSION

Data dilution: 983 pairs drowned the C3 signal. 0.878 heretic. More data ≠ better.

USECASE

Warning example of data dilution. Educational value only.

Full benchmark → | LoopKit →