reasoning-degeneration-dev/algo-sft-formal-logic-bottom-up
Updated • 12 • 1
Algorithmic SFT vs Distillation
10 LoRA adapters + 6 datasets. Algo template SFT vs QwQ distillation on Qwen2.5-1.5B-Instruct across 4 reasoning domains.
Note Formal Logic: Bottom-Up (algo) — 100% test, 92.6% OOD
reasoning-degeneration-dev/algo-sft-formal-logic-truth-table
Updated • 11Note Formal Logic: Truth Table (algo) — 100% test, 90.2% OOD
reasoning-degeneration-dev/algo-sft-formal-logic-distill-qwq
Updated • 12Note Formal Logic: QwQ Distill — 87.4% test, 71.2% OOD
reasoning-degeneration-dev/algo-sft-conlang-morphology-ordered-rules-d5d7
Updated • 12Note Conlang: Ordered Rules (algo) — 98.6% test, 94.2% OOD
reasoning-degeneration-dev/algo-sft-conlang-morphology-distill-qwq
Updated • 12Note Conlang: QwQ Distill — 40.4% test, 38.4% OOD
reasoning-degeneration-dev/algo-sft-cellular-automata-step-simulation-d5
Updated • 12Note CA: Step Sim d5 (algo) — 94.6% test, 72.0% OOD
reasoning-degeneration-dev/algo-sft-cellular-automata-distill-qwq
Updated • 12Note CA: QwQ Distill — 40.4% test, 22.4% OOD
reasoning-degeneration-dev/algo-sft-long-arithmetic-standard
Updated • 10Note Long Arith: Standard (algo) — 92.6% test, 0% OOD
reasoning-degeneration-dev/algo-sft-long-arithmetic-chunked
Updated • 10Note Long Arith: Chunked (algo) — 86.2% test, 0% OOD
reasoning-degeneration-dev/algo-sft-long-arithmetic-distill-qwq
Updated • 8Note Long Arith: QwQ Distill — 90.6% test, 6.8% OOD
reasoning-degeneration-dev/algorithmic-sft-full-eval-v3
Viewer • Updated • 50 • 21Note Aggregate eval results (v3)
reasoning-degeneration-dev/algorithmic-sft-training-data-v1
Viewer • Updated • 63k • 24Note Algo template training data (63K)
reasoning-degeneration-dev/algorithmic-sft-distillation-training-data-v1
Viewer • Updated • 24.1k • 35Note QwQ distillation training data (24K)
reasoning-degeneration-dev/algorithmic-sft-sharegpt-training-v1
Viewer • Updated • 82.9k • 37Note ShareGPT training data as-trained (83K)
reasoning-degeneration-dev/algorithmic-sft-eval-sets-v1
Viewer • Updated • 11k • 24Note Eval questions (11K)
reasoning-degeneration-dev/algorithmic-sft-training-configs-v1
Viewer • Updated • 17 • 27Note Training configs (17 YAMLs)