Hugging Face's logo

Collections

Discover the best community collections!

Collections trending this week
E2K cd82 checkpoints: sft_first_decl-cc2
Declarative-row trial 2 (qa+rule rows, anchor ratio 1.5), 3 SFT rollouts (iter 0-2); collapsed at its first RL rollout - negative result.
E2K cd82 checkpoints: sft_first_armc-cc
Arm C, 3 SFT rollouts: knowledge sentences inserted atop full base-model thinks, cite weight 1/8/32; knowledge NLL 3.05 -> 1.56; no recital yet.
E2K cd82 checkpoints: act_ctrl-cc
Matched control of IDD: same schedule, SFT rows = the base model's own turns without knowledge (iter 0-1), then 2 RL rollouts; stopped by decision.
E2K cd82 checkpoints: sft_first_idd-cc
IDD arm: 2 SFT rollouts on 303 teacher-recite harness rows (knowledge in the teacher's context), Adam reset, 7 RL rollouts (iter 2-8); +0.25.
E2K cd82 checkpoints: sft_first_decl-cc
Declarative-row trial 1 (text ending), 4 SFT rollouts (iter 0-3); first RL rollout collapsed into single-call episodes (44%) - negative result.
E2K cd82 checkpoints: sft_first_decl-cc4
Declarative chat rows (claim form), tool-call ending + anchor rows: 3 SFT rollouts (iter 0-2), 1 RL rollout (iter 3); no collapse, thinks shrink 3x.
E2K cd82 checkpoints: baseline-cc
Pure Dr.GRPO RL, no SFT, 12 rollouts x 32 episodes via the Claude Code harness (iter 0-11); shaped return -0.26 -> +0.08, 2/32 clears at rollout 10.
E2K cd82 checkpoints: sft_first_armc-cc6
Arm C continued from iter 2: 3 more SFT rollouts (knowledge NLL -> 0.42, iter 3-5) + 1 RL rollout (iter 6); first policy reciting knowledge verbatim.
E2K cd82 checkpoints: sft_first_decl-cc2
Declarative-row trial 2 (qa+rule rows, anchor ratio 1.5), 3 SFT rollouts (iter 0-2); collapsed at its first RL rollout - negative result.
E2K cd82 checkpoints: sft_first_decl-cc
Declarative-row trial 1 (text ending), 4 SFT rollouts (iter 0-3); first RL rollout collapsed into single-call episodes (44%) - negative result.
E2K cd82 checkpoints: sft_first_armc-cc
Arm C, 3 SFT rollouts: knowledge sentences inserted atop full base-model thinks, cite weight 1/8/32; knowledge NLL 3.05 -> 1.56; no recital yet.
E2K cd82 checkpoints: sft_first_decl-cc4
Declarative chat rows (claim form), tool-call ending + anchor rows: 3 SFT rollouts (iter 0-2), 1 RL rollout (iter 3); no collapse, thinks shrink 3x.
E2K cd82 checkpoints: act_ctrl-cc
Matched control of IDD: same schedule, SFT rows = the base model's own turns without knowledge (iter 0-1), then 2 RL rollouts; stopped by decision.
E2K cd82 checkpoints: baseline-cc
Pure Dr.GRPO RL, no SFT, 12 rollouts x 32 episodes via the Claude Code harness (iter 0-11); shaped return -0.26 -> +0.08, 2/32 clears at rollout 10.
E2K cd82 checkpoints: sft_first_idd-cc
IDD arm: 2 SFT rollouts on 303 teacher-recite harness rows (knowledge in the teacher's context), Adam reset, 7 RL rollouts (iter 2-8); +0.25.
E2K cd82 checkpoints: sft_first_armc-cc6
Arm C continued from iter 2: 3 more SFT rollouts (knowledge NLL -> 0.42, iter 3-5) + 1 RL rollout (iter 6); first policy reciting knowledge verbatim.