This repository archives the paper-critical Qwen2.5-7B-Instruct checkpoints from the FrontierDiag static-curriculum diagnostic. All schedules use the same 7,168-example multiset, exposure counts, constant learning rate, and supervised response-token budget. Only sample order changes.

Schedules and endpoints

random: fixed random interleaving.
easy-to-hard: Levels 1 to 5.
hard-to-easy: Levels 5 to 1.
Seed: 10.
Steps 112, 224, and 336 correspond to one, two, and three complete data exposures.

Every archived subfolder is a directly loadable Transformers checkpoint:

python

from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "huggingFacing/frontierdiag-qwen25-7b-curriculum-v1"
subfolder = "hard-to-easy/seed10/step112-pass1"
tokenizer = AutoTokenizer.from_pretrained(repo_id, subfolder=subfolder)
model = AutoModelForCausalLM.from_pretrained(repo_id, subfolder=subfolder)

These checkpoints are intended for diagnosis and reproducibility, not as general-purpose instruction models.