SFT of Qwen/Qwen3-8B on a 1000-row random mixed subset of ethanlshen/sera-subset (stage1 unresolved + stage2 resolved), trained with axolotl following the upstream SERA recipe.

See baselines/sera/README.md in the open-thoughts/OpenThoughts-Agent repo for the full reproduction details, hyperparameters, and iteration history (this is iteration i9, version v8).

Hyperparameters

learning_rate: 1e-5
batch_size: 32 (global; micro=1, grad_accum=1, dp=32)
num_epochs: 3
warmup_steps: 48
adam_beta1: 0.9, adam_beta2: 0.95
weight_decay: 0.01
sequence_len: 32768
chat_template: chatml
bf16, deepspeed zero3 (no CPU offload)

laion/sera-subset-mixed-1000-axolotl__Qwen3-8B-v8

sera-subset-mixed-1000-axolotl__Qwen3-8B-v8

Hyperparameters

laion/
sera-subset-mixed-1000-axolotl__Qwen3-8B-v8