shiyuxie/method_checkpoint

first_stop_match online-update checkpoints

LoRA checkpoints from a 40-update online-intervention run (OTMethod, first_stop_match), one directory per committed update.

updates	0000 (initialisation) .. 0040
adapter	PEFT LoRA, r=16, alpha=32, dropout=0.0, target_modules=[q_proj, v_proj]
per update	adapter_model.bin (7.0 MB), training.pt (22 MB: optimizer + RNG), adapter_config.json, checkpoint.json
total	123 files, 1.11 GiB

Integrity

Every file is uploaded byte-identical to the run output. checkpoint.json in each directory carries the sha256 of adapter_model.bin, training.pt and adapter_config.json; all 123 were re-verified against those records before upload and matched. Do not rewrite adapter_config.json in place - that breaks the recorded hash the training code checks.

Base model

adapter_config.json records base_model_name_or_path as the local filesystem path used by the run (the base model is the one this project refers to as M10, a 4B instruct model, revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a). Repoint it to your own copy before loading, for example:

from peft import PeftConfig, PeftModel
from transformers import AutoModelForCausalLM

config = PeftConfig.from_pretrained("shiyuxie/method_checkpoint", subfolder="0040")
config.base_model_name_or_path = "<your local or hub path to the M10 base>"
base = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path)
model = PeftModel.from_pretrained(base, "shiyuxie/method_checkpoint", subfolder="0040", config=config)

Result of this run

Neither primary metric improved. GPQA-Diamond, 198 questions, 4 samples each, paired-question bootstrap (10,000 draws):

• accuracy 76.89% -> 76.52%, -0.38 pp, 95% CI [-2.65, +1.77]
• thinking tokens 11,721 -> 12,039, +318, 95% CI [-178, +831]
• relative thinking-cost saving -2.71%, 95% CI [-7.10%, +1.50%]

Both intervals cross zero and the pre-registered criterion (accuracy up and cost down) is not met. The full analysis/ output of the run carries the paired table, metrics and figures.