mimo-7b-gdn-hybrid-400M-OPD
Stage-3 OPD @ 400M generated tokens — the H=16K MiMo control extended from
200M to 403.27M tokens (WSD schedule: flat 1e-5 to ~340M, 482-step linear decay
to 0). Snapshot final of run mimo-rugfix-h200x3-ext200-decay0-v1 (W&B
project mimo-gdn-linearization). Evals in full_eval/: GSM8K@1024 strict
0.58908 / flexible 0.67703; RULER-lite n=100 seed 0: niah_multikey@32768 0.82,
niah_single 1.00 at all lengths; AIME'24/'25 (think) included when complete.
Part of the MiMo second-model arm of a study converting full-attention models into GDN (gated DeltaNet) hybrids (uniform 1:4 retention) and recovering capability via staged distillation + on-policy distillation. Teacher/base: MiMo-7B-RL-0530.
NOTE: these checkpoints use the custom mimo_gdn architecture — register it
before loading (import src.models.mimo_gdn from the project repo, or use the
load_causal_lm helper in src/eval/sample_generations.py). Bare
AutoModelForCausalLM will not work. vLLM serving needs
src.models.vllm_mimo_gdn.register().
Sibling checkpoints: mimo-7b-gdn-hybrid-200M-OPD (base 200M control), and -500M/-600M-OPD from the H=32K branch when it completes.