q0-gsm8k-135m-mopd

q0 gsm8k distilled student

this is the selected single-model checkpoint from a small multi-teacher on-policy distillation experiment on GSM8K.

the student starts from HuggingFaceTB/SmolLM2-135M-Instruct at revision 12fd25f77366fa6b3b4b768ec3050bf629380bac. it generates its own responses, then learns with the uniform average forward KL from four frozen q0 snapshots. every teacher contributes 25%. learned mixture weights are used only to select the four teacher names and never enter the loss.

selection

training used 2,048 prompts for 32 steps. candidates were saved at steps 8, 16, 24, and 32, then compared on a fresh 512-example validation slice with eight sampled answers per question. step 8 had the best validation pass@8 at 5.47% and became opsd.pt.

official gsm8k test

all results use the complete 1,319-example test split.

the checkpoint matched the best baseline under greedy decoding, but its pass@k scores were lower. this is an exploratory negative result, not a reasoning improvement claim.

loading

opsd.pt is a bfloat16 PyTorch state dict. use the exact base revision before loading it.

import torch
from huggingface_hub import hf_hub_download
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "Pradheep1647/q0-gsm8k-135m-mopd"
base = "HuggingFaceTB/SmolLM2-135M-Instruct"
revision = "12fd25f77366fa6b3b4b768ec3050bf629380bac"

tokenizer = AutoTokenizer.from_pretrained(base, revision=revision)
model = AutoModelForCausalLM.from_pretrained(base, revision=revision, torch_dtype=torch.bfloat16)
checkpoint = hf_hub_download(repo_id=repo, filename="opsd.pt")
state = torch.load(checkpoint, map_location="cpu", weights_only=True)
model.load_state_dict(state)

selected checkpoint sha256: f9fcd1bd1df97e08d6b135fa1b276fcd9e8eba3dfb8bc3a7776876f95b782dc7

the repository also includes training metrics, validation results, official test results, full logs, plots, and the exact training and evaluation scripts.