reagent-lexsi/sft-aya32b-r16-v2
Updated • 60
Reflexive Agents: models
Every LoRA adapter for Aya Expanse 32B from Reflexive Agents, adopted or not, each with its result
Note Arm B, the healed model: 51.5% of τ²-bench test tasks (untuned: 14.5%)
reagent-lexsi/heal-trajectory-cycle-c1
Updated • 15Note Healed by the automated loop: dev 6 → 22 of 36, promoted by the gate
reagent-lexsi/heal-first-cycle-c1-rejected
Updated • 28Note First automated heal cycle, one turn per fix: findings −71%, rejected
reagent-lexsi/sft-aya32b-r16
Updated • 32Note SFT v1: dev 11% → 38%, rejected on airline
reagent-lexsi/sft-aya32b-r16-long
Updated • 25Note SFT SL: v1's data at 74 steps (ablation)
reagent-lexsi/sft-aya32b-r16-v3
Updated • 27Note SFT v3: + 20 retail passes; matched v2, not adopted
reagent-lexsi/sft-aya32b-r16-v4
Updated • 28Note SFT v4: + 29 on-policy fixes; matched v2, not adopted
reagent-lexsi/grpo-aya32b-v2
Updated • 28Note GRPO G: collapsed (0.5/30), a negative result
reagent-lexsi/grpo-heal-aya32b-v2
Updated • 27Note GRPO H: halved loops, rejected by the gate
reagent-lexsi/grpo-decision-aya32b-v2
Updated • 26Note GRPO H2: decision reward, 14.5/30 against v2's 17.0
reagent-lexsi/grpo-decision-h3-aya32b-v2
Updated • 38Note GRPO H3: H2 revised, not adopted