paper-gmts-token-select-1.5b

Recipe: recipes/papers/gmts-token-select · Collection: Papers, replicated

GRPO on GSM8K training on 20% of the batch's tokens, ranked by entropy (baseline) or by entropy times advantage (recipe). Both arms keep the same number of tokens, so the comparison isolates which tokens, not how many.

Result

Arm	pass@1	95% CI	pass@k	Steps	GPU min
Base, no training	0.36	[0.29, 0.43]	0.58	0	0
Baseline (top 20% by entropy)	0.49	[0.42, 0.57]	0.68	40	27.6
Recipe (top 20% by entropy x advantage)	0.18	[0.13, 0.23]	0.41	40	17.7

Recipe vs baseline: -0.310 [-0.383, -0.235] over 120 paired tasks. The recipe arm finished below the untrained model: mean completion length fell from 700 to 374 characters and the hack scan flagged the collapse. These weights are published as the record of that run, not as a model to use.

Arms in this repo

The root holds the arm the recipe README's headline number reports. Every other arm is a subfolder named after it. checkpoints/ never ships.

folder	arm
.	recipe arm: entropy x advantage selection, 2026-09-18 run (collapsed)
baseline	baseline arm: entropy selection, 2026-09-18 run

Load

from peft import PeftModel
from transformers import AutoModelForCausalLM

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")
model = PeftModel.from_pretrained(base, "while-ai/paper-gmts-token-select-1.5b")  # the headline arm
model = PeftModel.from_pretrained(base, "while-ai/paper-gmts-token-select-1.5b", subfolder="baseline")  # another arm

Reproduce

git clone https://github.com/whilehq/whileai-sdk && cd whileai-sdk/recipes/papers/gmts-token-select
python recipe.py

The recipe README pins the seed, the library versions and the GPU, and its Checks table says what the eval verified. Read the Learned section before quoting a number from this card.