andrewkang12345/policyINR-checkpoints

policyINR — Checkpoints

Companion checkpoint repository for andrewkang12345/policyINR — robust policy representation learning from offline data.

The folder layout mirrors the outputs/ tree of the source repo:

<domain>/<suite>/<run>/best.pt
                        last.pt

where <domain> ∈ {lichess, droid, fastf1, mujoco, syntheticgrf, dmlabseekavoid}.

Every per-run dir on the GitHub side carries the matching config.yaml, metrics.jsonl, summary.json, and eval.json — clone the repo and the identifiers line up 1-to-1 with the paths here.

Download example

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="andrewkang12345/policyINR-checkpoints",
    repo_type="model",
    allow_patterns=["lichess/2x_hk240/**/*.pt"],
    local_dir="outputs",
)

Domains

domain	what it is
lichess	Top-3 GM Lichess games — discrete UCI-move action
droid	DROID lowdim teleop — continuous action across collectors
fastf1	F1 stint telemetry — driver-as-policy
mujoco	Custom MuJoCo + Minari baselines + state/action-resampled suites
syntheticgrf	Synthetic Gaussian random-field policies (sanity / smoke)
dmlabseekavoid	DMLab seekavoid_arena_01 from RL Unplugged — discrete action

Models in each run name

<data>__<model>__<experiment>__s<seed>/

• cvae — bag-of-pairs CVAE
• inr_transformer_history_conditioned — INR-Transformer w/ history
• inr_diffusion_history_conditioned — INR-Diffusion w/ history
• inr_transformer_fitted_latent — per-unit fitted latent codes
• inr_transformer_infer_latent[_maml] — meta-learned per-unit latent (10-step inner adapt + early stopping)

License

MIT (mirrors the source repo).