agill32/precog-rm-checkpoints

Precog-RM checkpoints

Trained encoders from "Divergent Paths: Learning an Outcome-Predictive Representation of LLM Agent Trajectories" (Gill, Gogia, Guzmán), arXiv:2610.19614. Code and loaders: https://github.com/alexgill321/precog-rm (src/prm/train_precog.py:PrecogProtoNet.load).

checkpoints/sweep/precog_<pool>_s<seed>.pt      75 checkpoints of the main sweep (25 training pools x seeds 0-2):
                                                <benchmark>_<g|q|d combination> (g=GPT-5-mini, q=Qwen3-80B, d=DeepSeek-V3.2),
                                                two-benchmark and generalist pools as <b1>-<b2>[-<b3>]_all.
                                                10,000 episodes each; Tables 1-2, Figure 3, Appendix J.
checkpoints/design_study/src_<name>_s0.pt       single-seed design study on tau2/qwen (Appendices C, F; 4,000 episodes)
checkpoints/design_study/mix_<name>_s0.pt       the same ablations retrained on the all-cell pool
checkpoints/design_study/pool_<pool>_full_s0.pt epoch-resolved pool sweep (notebook study)
checkpoints/prefix/prefix_<pool>_<variant>_s0.pt prefix-trained encoders (Section 4.5, Figure 4)
checkpoints/grpo_controller.pt                  the controller deployed inside GRPO (Section 5); trained on
                                                tau2/qwen with scripts/run_prefix_verify.py (predates the final sweep protocol)

scripts/download.py --part checkpoints fetches this repo into $PRECOG_DATA/ckpts/ and re-folds it into the layout the code reads: ckpts/precog/, ckpts/ablations/, ckpts/prefix/ (with the controller as ckpts/prefix/grpo_controller_model_prefix.pt).

Architecture: linear projection of 3072-d step embeddings to d'=128, 2-layer transformer (4 heads, FFN 512, dropout 0.3), masked mean pooling, unsquared-Euclidean prototypes. Produced by scripts/precog_sweep.py, scripts/design_study.py --train, scripts/prefix_gating_eval.py --train.