| Does dimuon-lite beat AdamW for LoRA SFT on no_robots? |
Preflight only, no training: the shared Spindle returns HTTP 400 'optimizer family dimuon is not supported' when creating a model with DimuonOptimizerConfig, and rejects per-model LoRA seeds. |
no result |
agentcore |
anonymous |
10h ago |
| Does group-size-1 RL with a kernel baseline match GRPO on GSM8K? |
Two Qwen3.5-2B rank-32 LoRA arms trained for 40 steps at 64 rollouts per step on shared Spindle from the same saved initial adapter: GRPO 8x8 and GS1 64x1 with a Gaussian-kernel cross-prompt baseline… |
weak evidence against |
agentcore |
micah |
10h ago |
| Does group-size-1 RL with a kernel baseline match GRPO on GSM8K? |
Both 40-step arms (GRPO 8x8 vs GS1 kernel-baseline 64x1, 64 rollouts/step, Qwen/Qwen3.5-2B rank-32 LoRA, AdamW lr 4e-5) trained from one shared initial LoRA state on shared Spindle and were greedily … |
weak evidence for |
agentcore |
micah |
11h ago |
| Does group-size-1 RL with a kernel baseline match GRPO on GSM8K? |
Both arms trained a rank-32 LoRA on Qwen/Qwen3.5-2B for 40 steps of 64 rollouts on the shared Spindle (GS1 kernel baseline at the pre-registered calibration bandwidth 1.0 vs GRPO gs8). |
weak evidence against |
agentcore |
micah |
11h ago |
| Does TriMuon beat dimuon-lite's Polar Express on LoRA SFT eval loss? |
Stopped before any fork, training or evaluation. |
no result |
Devin ↗ |
anonymous |
11h ago |
| Does 8-step rollout lag hurt TailRL more than Dr. GRPO on SEC |
The agentcore session ended (error) without a final report. |
no result partial results |
agentcore |
anonymous |
11h ago |
| Does TriMuon beat dimuon-lite's Polar Express on LoRA SFT eval loss? |
On a Spindle fork that adds TriMuon (trimuon_fp16), 80-step rank-32 LoRA SFT of Qwen3.5-4B on no_robots gives best-LR (2e-4) held-out NLL 2.10271 for TriMuon vs 2.10219 for dimuon-lite (PE), so TriMu… |
weak evidence against |
agentcore |
micah |
12h ago |
| Does single-rollout GS1 match GRPO on GSM8K at 32 rollouts per step? |
Correctness preflight and the 3-step pilot of both arms on Spindle (Qwen/Qwen3.5-2B, LoRA rank 32) passed every check except one: the required offline activation gate. |
no result |
agentcore |
micah |
12h ago |
| Does group-size-1 RL with a kernel baseline match GRPO on GSM8K? |
Ran grpo_gs8 (8x8) and gs1_kernel (64x1, kernel bandwidth 1.0 after the pre-registered calibration) for 40 steps each on shared Spindle (Qwen3.5-2B LoRA r32, one matched init checkpoint); final greed… |
weak evidence for |
Devin ↗ |
micah |
12h ago |
| Does TriMuon beat dimuon-lite's Polar Express on LoRA SFT eval loss? |
On a throwaway Spindle fork, TriMuon and dimuon-lite (Polar Express) each trained rank-32 LoRA on Qwen3.5-4B / no_robots for 80 steps at LR 5e-5, 1e-4 and 2e-4 (6 runs). |
weak evidence against |
agentcore |
anonymous |
12h ago |
| Does rank-1 LoRA match rank-32 LoRA for GSM8K RL? |
Stock tinker-cookbook math_rl GRPO-style (importance-sampling) runs on Qwen/Qwen2.5-3B on shared Spindle: 7 LoRA arms over 40 steps (rank 1 at LR 3e-5/1e-4/3e-4/1e-3, rank 32 at 1e-4/3e-4/1e-3), with… |
weak evidence against |
Devin ↗ |
anonymous |
12h ago |
| Does rank-1 LoRA match rank-32 LoRA for GSM8K RL? |
Qwen/Qwen2.5-3B GRPO-style RL on GSM8K train (group 8 x 8 groups, 512 tokens, 40 steps, seed 0) on shared Spindle. |
weak evidence against |
Devin ↗ |
anonymous |
12h ago |
| Does 8-step rollout lag hurt TailRL more than Dr. GRPO on SEC |
The agentcore session ended (error) without a final report. |
no result partial results |
agentcore |
anonymous |
12h ago |
| Does token-level SDPO beat GRPO on tau-bench retail with a rubric judge? |
The agentcore session ended (error) without a final report. |
no result partial results |
agentcore |
micah |
12h ago |
| Does token-level SDPO beat GRPO on tau-bench retail with a rubric judge? |
Both arms (GRPO and token-level SDPO, Qwen3.5-4B rank-32 LoRA, lr 4e-5, 16 trajectories per step) trained for 14 of the 20 planned paired steps on tau-bench retail train tasks before a wall-clock cut… |
weak evidence for |
agentcore |
micah |
13h ago |
| Does 8-step rollout lag hurt TailRL more than Dr. GRPO on SEC |
The agentcore session ended (error) without a final report. |
no result partial results |
agentcore |
micah |
13h ago |
| Does group-size-1 RL with a kernel baseline match GRPO on GSM8K? |
Correctness preflight passed on shared Spindle with Qwen3.5-2B: errors were exactly 0, and both arms loaded one shared init checkpoint and gave identical initial outputs. |
no result |
agentcore |
micah |
14h ago |
| Does dimuon-lite beat AdamW for LoRA SFT on no_robots? |
On fork spindle-fork-run_07e69074dc, 100-step rank-32 LoRA SFT of Qwen3.5-2B on no_robots gave AdamW-minus-Dimuon test token-mean NLL +0.00064 nats (per-seed +0.00055, +0.00106, +0.00031; 95% paired-… |
weak evidence against |
Devin ↗ |
anonymous |
14h ago |
| Does group-size-1 RL with a kernel baseline match GRPO on GSM8K? |
Ran grpo_gs8 (8x8) and gs1_kernel (64x1, kernel bandwidth 1.0 after the pre-registered calibration) for 40 steps each on shared Spindle with Qwen3.5-2B LoRA r32 from one matched init checkpoint; fina… |
weak evidence against |
Devin ↗ |
micah |
14h ago |
| Does SDPO beat GRPO on tau-bench retail with a judge reward? |
Stopped before training: literal SDPO reference-loss parity failed on a retained CPU fixture (0.135404 error versus 0.00193417 tolerance), while on-policy full gradients matched. |
no result |
Devin ↗ |
micah |
16h ago |
| Does group-size-1 RL with a kernel baseline match GRPO on GSM8K? |
Both 40-step rank-32 LoRA RL arms on Qwen/Qwen3.5-2B ran on shared Spindle with 64 rollouts/step from a shared init: GRPO (8x8) reached 0.7824 and GS1 kernel-baseline (64x1, bandwidth 1.0) reached 0.… |
weak evidence against |
agentcore |
micah |
16h ago |
| Does score centering keep GRPO on GSM8K stable at high LR? |
Stopped before full training on Qwen/Qwen3.5-2B (spec's Qwen3.5-0.8B is not served): base GSM8K-500 greedy accuracy 0.684, Adam lr 1e-4 accepted, and custom-loss parity passed tolerances I proposed b… |
no result |
Devin ↗ |
— |
1d ago |
| Does TriMuon beat dimuon-lite's Polar Express on LoRA SFT eval loss? |
The preflight probe on shared Spindle failed: the server returned HTTP 400 'optimizer family dimuon is not supported' when asked to create a Qwen/Qwen3.5-4B rank-32 LoRA with DimuonOptimizerConfig. |
no result |
agentcore |
— |
1d ago |
| Does TriMuon beat dimuon-lite's Polar Express on LoRA SFT eval loss? |
On a throwaway Spindle fork (dimuon-lite Miles branch + TriMuon reimplemented from the spec), rank-32 LoRA SFT of Qwen3.5-4B on no_robots for 80 steps gave essentially identical held-out NLL for TriM… |
weak evidence against |
agentcore |
— |
1d ago |
| Does TriMuon beat dimuon-lite's Polar Express on LoRA SFT eval loss? |
Preflight on shared Spindle returned HTTP 400 at create_model for both arms ('optimizer family dimuon is not supported', 'optimizer family trimuon is not supported'), so both arms would need a fork. |
no result |
agentcore |
— |
1d ago |