cm00cm/Qwen3.5-35B-A3B-DAPO-RLVR-teacher

Qwen3.5-35B-A3B — DAPO RLVR Teacher (Phase 1)

RLVR (GRPO) fine-tune of Qwen3.5-35B-A3B on DAPO-math-17k, trained with miles on a single 8×H200 node. This is the Phase-1 teacher for an on-policy self-distillation pipeline (see miles example examples/on_policy_distillation/qwen3_5_35b_selfdistill).

What changed vs the base

Held-out DAPO-math (512 disjoint problems, eval @ 24k cap, temperature 0.6):

	accuracy	mean response length
base	0.828	14,070
this checkpoint	0.887	6,248

The RLVR step (lr 1e-5, GRPO + format-agnostic boxed/Answer correctness reward, 24k rollout cap) makes the model both more accurate and ~2× more concise.

Details

• Parallelism: TP2 / PP1 / CP2 / EP8 / ETP1 (single node, world=8), --colocate, optimizer CPU offload.
• Reward: format-agnostic correctness (accepts \boxed{} or Answer:).
• Weights only (optimizer state stripped). Standard transformers load.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("cm00cm/Qwen3.5-35B-A3B-DAPO-RLVR-teacher")
model = AutoModelForCausalLM.from_pretrained(
    "cm00cm/Qwen3.5-35B-A3B-DAPO-RLVR-teacher", torch_dtype="bfloat16", device_map="auto")