| Does 8-step rollout lag hurt TailRL more than Dr. GRPO on SEC |
Qwen3.5-2B base GSM8K evaluation measured pass@1=0.6015 and pass@8=0.878 on 500 problems, and 17 TailRL lag-0 optimizer updates completed with stock Tinker. |
no result |
Devin ↗ |
— |
2d ago |
| Does dimuon-lite beat AdamW for LoRA SFT on no_robots? |
A Spindle fork (app spindle-dimuon-fork, Qwen/Qwen3.5-2B because 0.8B is not served) accepted and ran dimuon-lite optim_steps, but the 6-run LR sweep was stopped at step 5-6 of 100 because the measur… |
no result |
Devin ↗ |
— |
2d ago |
| Does DPPO's divergence mask beat GRPO clipping on GSM8K? |
On Spindle, 100-step rank-32 LoRA RL on GSM8K with Qwen/Qwen3.5-2B (substituted because Qwen3.5-0.8B is not served) compared the stock cookbook GRPO importance-sampling loss against a DPPO-Binary-TV … |
no result |
Devin ↗ |
— |
2d ago |
| Does eight-token score centering beat importance-sampled REINFORCE under stale rollouts? |
Stock Tinker custom-loss and sampler probes worked on the permitted Qwen3.5-2B fallback, but the paired GSM8K run and an initialization retry failed because Spindle reported trainer capacity exhausted. |
no result |
Devin ↗ |
— |
2d ago |
| Does GS1 one-rollout REINFORCE match GRPO on GSM8K? |
On Spindle with Qwen/Qwen3.5-2B (0.8B is not served), rank-32 LoRA, 100 steps x 128 rollouts, the base model scored 69.98% on all 1319 GSM8K test problems; GRPO (16x8) scored 82.26% and 77.63% on see… |
no result |
Devin ↗ |
— |
2d ago |
| Does dimuon-lite beat AdamW for LoRA SFT on no_robots? |
This can't run on Spindle: the shared server rejects every dimuon-lite optim_step with HTTP 400 "optimizer family dimuon is not supported; optim_step accepts only adam_params", so only the AdamW arm … |
no result |
Devin ↗ |
— |
2d ago |
| Does DPPO's divergence mask beat GRPO clipping on GSM8K? |
The Devin session ended (finished) without reporting a result. |
no result |
Devin ↗ |
— |
2d ago |
| Does eight-token score centering beat importance-sampled REINFORCE under stale rollouts? |
RL on openai/gsm8k with Qwen/Qwen3.5-2B (LoRA rank 32, lr 4e-05, max_tokens 512), eval on 256 held-out problems with greedy decoding. |
no result |
runner |
— |
2d ago |
| Does LoRA SFT make Qwen3.5-4B a better PubMedQA grader? |
Train loss 0.803 -> 0.018 over 150 steps on 9600 examples of qiaojin/PubMedQA (Qwen/Qwen3.5-4B, LoRA rank 32, adam lr 0.0001). |
no result |
runner |
— |
2d ago |
| Does dimuon-lite beat AdamW for LoRA SFT on no_robots? |
The run of 'dimuon-lite vs AdamW for LoRA SFT on no_robots (Qwen3.5-0.8B, rank 32, 100 steps)' on Qwen/Qwen3.5-0.8B did not produce a result (exit code 1) after 14.9 minutes. |
no result |
runner |
— |
2d ago |
| Does GS1 one-rollout REINFORCE match GRPO on GSM8K? |
The run of 'GS1 (group-size-1, similarity-weighted LOO batch baseline) vs GRPO on GSM8K at equal rollout budget' on Qwen/Qwen3.5-0.8B did not produce a result after 0.0 minutes. |
no result |
runner |
— |
2d ago |