Entry point: 5 datasets (trajectories, verified knowledge, RL runs) + one checkpoint per RL run; each run has its own series collection.
Ziyan Wang
ziyan98
Mi6paulino's profile picture
webbrain-one-485467's profile picture
webbrain-one-730947's profile picture·
AI & ML interests
None yet
Recent Activity
updated a model 14 days ago
ziyan98/E2K-cd82-Qwen3.8-27B-sft_inter_decl-cc2-iter0 updated a model 14 days ago
ziyan98/E2K-cd82-Qwen3.8-27B-sft_inter_decl-cc2-iter1 updated a model 14 days ago
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_decl-cc2-iter0Organizations
First IDD attempt (iter 2-6): SFT Adam moments carried into RL froze the RL step (~10x smaller moves); relaunched with Adam reset.
-
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_idd-cc-iter6-frozen-adam
Image-Text-to-Text • 28B • Updated • 50 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_idd-cc-iter5-frozen-adam
Image-Text-to-Text • 28B • Updated • 43 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_idd-cc-iter4-frozen-adam
Image-Text-to-Text • 28B • Updated • 44 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_idd-cc-iter3-frozen-adam
Image-Text-to-Text • 28B • Updated • 39
2026-09-12 atom arm: SFT-only checkpoint (iter 3) on 888 QA + 178 statement rows from the 888-atom set; cancelled that day for the claim-form ablation
2026-09-12 claim-form arm, SFT rows in every RL batch, in-process MCP client (iter 3/7/11); collapsed, superseded.
2026-09-12 claim-form arm, SFT-first schedule with the in-process MCP client (pre-harness RL), iter 3-19; collapsed, superseded.
-
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first-claimform-iter19
Image-Text-to-Text • 28B • Updated • 41 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first-claimform-iter15
Image-Text-to-Text • 28B • Updated • 40 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first-claimform-iter11
Image-Text-to-Text • 28B • Updated • 34 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first-claimform-iter7
Image-Text-to-Text • 28B • Updated • 38
First harness SFT arm, 391 claim-form QA rows (old JSON prompt), iter 3/7/11/15: the policy answered with one text turn and never acted (collapse).
-
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first-cc-iter15
Image-Text-to-Text • 28B • Updated • 38 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first-cc-iter11
Image-Text-to-Text • 28B • Updated • 46 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first-cc-iter7
Image-Text-to-Text • 28B • Updated • 34 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first-cc-iter3
Image-Text-to-Text • 28B • Updated • 42
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.
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.
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.
-
ziyan98/E2K-cd82-Qwen3.8-27B-act_ctrl-cc-iter3
Image-Text-to-Text • 28B • Updated • 39 -
ziyan98/E2K-cd82-Qwen3.8-27B-act_ctrl-cc-iter2
Image-Text-to-Text • 28B • Updated • 38 -
ziyan98/E2K-cd82-Qwen3.8-27B-act_ctrl-cc-iter1
Image-Text-to-Text • 28B • Updated • 48 -
ziyan98/E2K-cd82-Qwen3.8-27B-act_ctrl-cc-iter0
Image-Text-to-Text • 28B • Updated • 37
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.
-
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_idd-cc-iter8
Image-Text-to-Text • 28B • Updated • 40 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_idd-cc-iter7
Image-Text-to-Text • 28B • Updated • 43 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_idd-cc-iter6
Image-Text-to-Text • 28B • Updated • 42 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_idd-cc-iter5
Image-Text-to-Text • 28B • Updated • 43
First act-ctrl attempt (iter 2-6) with frozen Adam moments carried into RL; relaunched with an Adam reset as act_ctrl-cc.
-
ziyan98/E2K-cd82-Qwen3.8-27B-act_ctrl-cc-iter6-frozen-adam
Image-Text-to-Text • 28B • Updated • 48 -
ziyan98/E2K-cd82-Qwen3.8-27B-act_ctrl-cc-iter5-frozen-adam
Image-Text-to-Text • 28B • Updated • 40 -
ziyan98/E2K-cd82-Qwen3.8-27B-act_ctrl-cc-iter4-frozen-adam
Image-Text-to-Text • 28B • Updated • 37 -
ziyan98/E2K-cd82-Qwen3.8-27B-act_ctrl-cc-iter3-frozen-adam
Image-Text-to-Text • 28B • Updated • 34
2026-09-11 pilot of the SFT-first schedule on a hand-curated 36-atom level-1 set (44 rows), in-process client; iter 3.
Pre-harness pure-RL baseline (in-process MCP client, 2026-09-10/11, iter 3/7): reward 0 in every episode, zero gradient, weights equal the base.
2026-09-12 claim-form arm, SFT interleaved between RL rollouts, in-process MCP client (iter 3/7/11); collapsed, superseded.
-
ziyan98/E2K-cd82-Qwen3.8-27B-sft_interleaved-claimform-iter11
Image-Text-to-Text • 28B • Updated • 41 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_interleaved-claimform-iter7
Image-Text-to-Text • 28B • Updated • 48 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_interleaved-claimform-iter3
Image-Text-to-Text • 28B • Updated • 43
First harness joint arm: 391 QA + 79 statement rows mixed into every RL batch (iter 3/7/11); play collapsed into no-op single calls from rollout 2.
Declarative-row trial 3, interleaved schedule with the trial-2 rows, 2 SFT rollouts (iter 0-1); collapsed at its first RL rollout - negative result.
Declarative-row trial 1 (text ending), 4 SFT rollouts (iter 0-3); first RL rollout collapsed into single-call episodes (44%) - negative result.
-
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_decl-cc-iter3
Image-Text-to-Text • 28B • Updated • 47 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_decl-cc-iter2
Image-Text-to-Text • 28B • Updated • 37 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_decl-cc-iter1
Image-Text-to-Text • 28B • Updated • 39 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_decl-cc-iter0
Image-Text-to-Text • 28B • Updated • 37
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.
-
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_decl-cc4-iter3
Image-Text-to-Text • 28B • Updated • 44 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_decl-cc4-iter2
Image-Text-to-Text • 28B • Updated • 37 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_decl-cc4-iter1
Image-Text-to-Text • 28B • Updated • 40 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_decl-cc4-iter0
Image-Text-to-Text • 28B • Updated • 39
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.
-
ziyan98/E2K-cd82-Qwen3.8-27B-baseline-cc-iter11
Image-Text-to-Text • 28B • Updated • 42 -
ziyan98/E2K-cd82-Qwen3.8-27B-baseline-cc-iter10
Image-Text-to-Text • 28B • Updated • 44 -
ziyan98/E2K-cd82-Qwen3.8-27B-baseline-cc-iter9
Image-Text-to-Text • 28B • Updated • 42 -
ziyan98/E2K-cd82-Qwen3.8-27B-baseline-cc-iter8
Image-Text-to-Text • 28B • Updated • 43
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.
-
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_armc-cc6-iter6
Image-Text-to-Text • 28B • Updated • 38 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_armc-cc6-iter5
Image-Text-to-Text • 28B • Updated • 33 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_armc-cc6-iter4
Image-Text-to-Text • 28B • Updated • 35 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_armc-cc6-iter3
Image-Text-to-Text • 28B • Updated • 44
Entry point: 5 datasets (trajectories, verified knowledge, RL runs) + one checkpoint per RL run; each run has its own series collection.
First act-ctrl attempt (iter 2-6) with frozen Adam moments carried into RL; relaunched with an Adam reset as act_ctrl-cc.
-
ziyan98/E2K-cd82-Qwen3.8-27B-act_ctrl-cc-iter6-frozen-adam
Image-Text-to-Text • 28B • Updated • 48 -
ziyan98/E2K-cd82-Qwen3.8-27B-act_ctrl-cc-iter5-frozen-adam
Image-Text-to-Text • 28B • Updated • 40 -
ziyan98/E2K-cd82-Qwen3.8-27B-act_ctrl-cc-iter4-frozen-adam
Image-Text-to-Text • 28B • Updated • 37 -
ziyan98/E2K-cd82-Qwen3.8-27B-act_ctrl-cc-iter3-frozen-adam
Image-Text-to-Text • 28B • Updated • 34
First IDD attempt (iter 2-6): SFT Adam moments carried into RL froze the RL step (~10x smaller moves); relaunched with Adam reset.
-
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_idd-cc-iter6-frozen-adam
Image-Text-to-Text • 28B • Updated • 50 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_idd-cc-iter5-frozen-adam
Image-Text-to-Text • 28B • Updated • 43 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_idd-cc-iter4-frozen-adam
Image-Text-to-Text • 28B • Updated • 44 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_idd-cc-iter3-frozen-adam
Image-Text-to-Text • 28B • Updated • 39
2026-09-11 pilot of the SFT-first schedule on a hand-curated 36-atom level-1 set (44 rows), in-process client; iter 3.
2026-09-12 atom arm: SFT-only checkpoint (iter 3) on 888 QA + 178 statement rows from the 888-atom set; cancelled that day for the claim-form ablation
Pre-harness pure-RL baseline (in-process MCP client, 2026-09-10/11, iter 3/7): reward 0 in every episode, zero gradient, weights equal the base.
2026-09-12 claim-form arm, SFT rows in every RL batch, in-process MCP client (iter 3/7/11); collapsed, superseded.
2026-09-12 claim-form arm, SFT interleaved between RL rollouts, in-process MCP client (iter 3/7/11); collapsed, superseded.
-
ziyan98/E2K-cd82-Qwen3.8-27B-sft_interleaved-claimform-iter11
Image-Text-to-Text • 28B • Updated • 41 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_interleaved-claimform-iter7
Image-Text-to-Text • 28B • Updated • 48 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_interleaved-claimform-iter3
Image-Text-to-Text • 28B • Updated • 43
2026-09-12 claim-form arm, SFT-first schedule with the in-process MCP client (pre-harness RL), iter 3-19; collapsed, superseded.
-
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first-claimform-iter19
Image-Text-to-Text • 28B • Updated • 41 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first-claimform-iter15
Image-Text-to-Text • 28B • Updated • 40 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first-claimform-iter11
Image-Text-to-Text • 28B • Updated • 34 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first-claimform-iter7
Image-Text-to-Text • 28B • Updated • 38
First harness joint arm: 391 QA + 79 statement rows mixed into every RL batch (iter 3/7/11); play collapsed into no-op single calls from rollout 2.
First harness SFT arm, 391 claim-form QA rows (old JSON prompt), iter 3/7/11/15: the policy answered with one text turn and never acted (collapse).
-
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first-cc-iter15
Image-Text-to-Text • 28B • Updated • 38 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first-cc-iter11
Image-Text-to-Text • 28B • Updated • 46 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first-cc-iter7
Image-Text-to-Text • 28B • Updated • 34 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first-cc-iter3
Image-Text-to-Text • 28B • Updated • 42
Declarative-row trial 3, interleaved schedule with the trial-2 rows, 2 SFT rollouts (iter 0-1); collapsed at its first RL rollout - negative result.
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.
Declarative-row trial 1 (text ending), 4 SFT rollouts (iter 0-3); first RL rollout collapsed into single-call episodes (44%) - negative result.
-
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_decl-cc-iter3
Image-Text-to-Text • 28B • Updated • 47 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_decl-cc-iter2
Image-Text-to-Text • 28B • Updated • 37 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_decl-cc-iter1
Image-Text-to-Text • 28B • Updated • 39 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_decl-cc-iter0
Image-Text-to-Text • 28B • Updated • 37
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.
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.
-
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_decl-cc4-iter3
Image-Text-to-Text • 28B • Updated • 44 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_decl-cc4-iter2
Image-Text-to-Text • 28B • Updated • 37 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_decl-cc4-iter1
Image-Text-to-Text • 28B • Updated • 40 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_decl-cc4-iter0
Image-Text-to-Text • 28B • Updated • 39
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.
-
ziyan98/E2K-cd82-Qwen3.8-27B-act_ctrl-cc-iter3
Image-Text-to-Text • 28B • Updated • 39 -
ziyan98/E2K-cd82-Qwen3.8-27B-act_ctrl-cc-iter2
Image-Text-to-Text • 28B • Updated • 38 -
ziyan98/E2K-cd82-Qwen3.8-27B-act_ctrl-cc-iter1
Image-Text-to-Text • 28B • Updated • 48 -
ziyan98/E2K-cd82-Qwen3.8-27B-act_ctrl-cc-iter0
Image-Text-to-Text • 28B • Updated • 37
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.
-
ziyan98/E2K-cd82-Qwen3.8-27B-baseline-cc-iter11
Image-Text-to-Text • 28B • Updated • 42 -
ziyan98/E2K-cd82-Qwen3.8-27B-baseline-cc-iter10
Image-Text-to-Text • 28B • Updated • 44 -
ziyan98/E2K-cd82-Qwen3.8-27B-baseline-cc-iter9
Image-Text-to-Text • 28B • Updated • 42 -
ziyan98/E2K-cd82-Qwen3.8-27B-baseline-cc-iter8
Image-Text-to-Text • 28B • Updated • 43
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.
-
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_idd-cc-iter8
Image-Text-to-Text • 28B • Updated • 40 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_idd-cc-iter7
Image-Text-to-Text • 28B • Updated • 43 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_idd-cc-iter6
Image-Text-to-Text • 28B • Updated • 42 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_idd-cc-iter5
Image-Text-to-Text • 28B • Updated • 43
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.
-
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_armc-cc6-iter6
Image-Text-to-Text • 28B • Updated • 38 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_armc-cc6-iter5
Image-Text-to-Text • 28B • Updated • 33 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_armc-cc6-iter4
Image-Text-to-Text • 28B • Updated • 35 -
ziyan98/E2K-cd82-Qwen3.8-27B-sft_first_armc-cc6-iter3
Image-Text-to-Text • 28B • Updated • 44