aviralku/mr9b-layeraudit-8node-215

mr9b meta-reasoning RL — layer-audit — 8-node/215 checkpoints

Mid-training research checkpoints from a fully-async GRPO run with the pipelined frozen-E layer audit enabled (meta_reasoning.layer_audit.enable=true).

• Base: Qwen3.5 (~9B hybrid, model_type: qwen3_5, 32 layers)
• Method: fully-async GRPO meta-reasoning RL (VERL fully_async_policy)
• Layer audit: after each completed reasoning layer, a frozen executor audits it (classify explorations as new/refinement/verification/duplicate, diagnose tunnel vision + gaps) and injects one-step control feedback into the next MR step.
• Topology 215: 2 trainer / 1 rollout / 5 frozen-E nodes (8 nodes, TP8/PP1/DP2)
• Experiment: mr9b_layeraudit_8node_215_20260718_231300
• W&B project: meta_reason_rl_h100_5node

Contents

Each subfolder is a standalone HF model (weights + tokenizer + chat template):

• global_step_68/
• global_step_69/

from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained("aviralku/mr9b-layeraudit-8node-215", subfolder="global_step_69")
t = AutoTokenizer.from_pretrained("aviralku/mr9b-layeraudit-8node-215", subfolder="global_step_69")

Private research checkpoints; not an official release.