Plain GRPO. entropy_coeff=0 (the framework default of 1e-3 is explicitly disabled so "vanilla" means vanilla), symmetric PPO clip 0.2, rollout temperature 1.0.

Do not apply a chat template

Trained on raw prompt text. verl's RLHFDataset has apply_chat_template=False and it was never enabled. Applying Qwen2.5-Math's chat template at inference creates a train/eval mismatch measured at roughly 19 points of pass@1 on a sibling task.

python

from vllm import LLM, SamplingParams
llm = LLM(model="sandeep123/sqa-grpo-vanilla-step600", dtype="bfloat16", max_model_len=1536)
params = SamplingParams(n=6, temperature=1.0, top_p=1.0, top_k=-1, max_tokens=1024)
out = llm.generate([prompt_text], sampling_params=params)   # raw string, not llm.chat()

Validation metrics at this checkpoint

metric	value
pass@1	0.7949
pass@6	0.9727
step	600

Answer extraction (pre-registered). An answer is the content of the final \boxed{}; if absent, the last standalone A-E token. Responses with no extractable answer are scored incorrect, and all K rollouts stay in the denominator. This is ScienceQA's answer-choice accuracy, reported as "sampled answer accuracy (pass@1)".

Validation uses 256 held-out prompts, K=6, temperature 1.0, seed 42 -- pinned in code so every arm, including the temperature-1.2 arm, is scored under identical decoding.

Settings (identical across all baseline arms)


dataset	ScienceQA (`scienceqa_boxfix`)
epochs / steps	25 / 1250
batch / rollouts	128 prompts, K=6
learning rate	1e-6 constant
KL (in-reward)	0.01
max prompt / response	512 / 1024 tokens
format reward	0.03, constant, no decay
seed	42

Arm-specific: entropy_coeff=0.0, clip=0.2/0.2, rollout temperature=1.0.

Note on checkpoint selection

Quality-optimal and diversity-optimal checkpoints differ substantially: for these runs the best-pass@1 checkpoint lands near step 1000-1200 while the best-pass@6 checkpoint is near step 200-500. Both are published so that diversity results are not reported from a checkpoint chosen purely for accuracy.