SFT Cold Start Subsampling
H100 cold-start SFT reproduction: base, six checkpoints and original 10-05 reference. Subsampling findings are provisional.
Text Generation • 67B • Updated • 156Note Base initializer.
open-athena/Grug-67B-A2B-Datakit-SFT-Repro-262K-Step222-2026.10.25
Text Generation • 67B • Updated • 54Note Reproduction checkpoint; corrected EOS generation config. Eval requires explicit thinking. Results provisional.
open-athena/Grug-67B-A2B-Datakit-SFT-Repro-262K-Step471-2026.10.25
Text Generation • 67B • Updated • 39Note Reproduction checkpoint; corrected EOS generation config. Eval requires explicit thinking. Results provisional.
open-athena/Grug-67B-A2B-Datakit-SFT-Repro-262K-Step1024-2026.10.28
Text Generation • 67B • Updated • 50Note Reproduction checkpoint; corrected EOS generation config. Eval requires explicit thinking. Results provisional.
open-athena/Grug-67B-A2B-Datakit-SFT-Repro-262K-Step1536-2026.10.28
Text Generation • 67B • Updated • 2Note Reproduction checkpoint; corrected EOS generation config. Eval requires explicit thinking. Results provisional.
open-athena/Grug-67B-A2B-Datakit-SFT-Repro-262K-Step2048-2026.10.29
Text Generation • 67B • Updated • 41Note Reproduction checkpoint; corrected EOS generation config. Eval requires explicit thinking. Results provisional.
open-athena/Grug-67B-A2B-Datakit-SFT-Repro-262K-Step2338-2026.10.29
Text Generation • 67B • Updated • 1Note Reproduction checkpoint; corrected EOS generation config. Eval requires explicit thinking. Results provisional.
open-athena/Grug-67B-A2B-Datakit-SFT-262K-2026.10.05
Text Generation • 67B • Updated • 1.56kNote Original 10-05 reference; separate full cold-start run.