Part of the Meta-Harness v1b study — co-evolution GRPO (verl) where a shared-weight policy plays two roles: TA (grader / solver, single-turn) and Proposer (harness/scaffold editor via multi-turn SEARCH/REPLACE). Question: does co-evolving the policy and its evaluation harness help, and through which lever (policy vs scaffold)?


Domain	Math (symbolic)
Architecture	Qwen3_5ForCausalLM — 32 layers, hidden 2560
Precision	bf16 (merged from FSDP via `verl.model_merger`)
Notes	No-reclaim, sequence-parallel=2 long-context, symbolic math. (Qwen3.5-4B)

Collection: https://huggingface.co/collections/hyunseoki/meta-harness-v1b-policy-x-harness-co-evolution-6a5c9877f74920e8922808bc

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("hyunseoki/mh-v1b-math-4b-sym-sp2")
model = AutoModelForCausalLM.from_pretrained("hyunseoki/mh-v1b-math-4b-sym-sp2", torch_dtype="bfloat16")