Tiny-RL refined task checkpoints
SFT and RL checkpoints for the V3 pretraining comparison: five model sizes, 10/30 pretraining tokens per parameter, five mixes, and 19 tasks. The archive fills as runs finish; missing cases are still pending.
Each run uses 16,777,216 SFT input tokens followed by 65,536 RL training episodes from its exact final SFT state. Environment difficulty stays fixed across models.
Files are under cases/<protocol>/<model-and-task>/:
sft.pt, rl.pt, and JSON configuration, ancestry, timing, and SHA256 metadata.
The .pt files contain config and a PyTorch model state dictionary.
from huggingface_hub import hf_hub_download
import torch
path = hf_hub_download(
"davidheineman/tiny-rl-refined-v3-model-sweep",
"cases/<protocol>/<model-and-task>/rl.pt",
)
checkpoint = torch.load(path, map_location="cpu", weights_only=True)
weights = checkpoint["model"]
Raw task data, evaluation prompts and saved generations are not part of this archive.