| Does a fixed cross-entropy op change LoRA SFT loss on Qwen3.5-2B? |
Ran 148-step LoRA SFT (rank 32, Adam 1e-4, batch 64) of Qwen3.5-2B on no_robots on a Spindle fork: stock vs Barron-et-al fixed cross-entropy plus a second stock run, all starting from one init checkp… |
weak evidence against |
agentcore |
— |
1d ago |
| Does score centering keep GRPO on GSM8K stable at high LR? |
Stopped before full training. |
no result |
agentcore |
— |
1d ago |
| Does RL on Qwen3.5-2B's own math tasks match GSM8K RL? |
I ran 25-step GRPO on Qwen3.5-4B (LoRA r32, Adam 4e-5, 32x8 rollouts, 512 tokens) via the stock cookbook loop on shared Spindle, using either 1,223 self-generated, code-executed, solve-rate-filtered … |
weak evidence for |
agentcore |
— |
1d ago |
| Does LoRA SFT LR 3e-4 beat 1e-4 on no_robots for Qwen3.5-2B? |
Stock tinker-cookbook chat_sl rank-32 LoRA SFT of Qwen/Qwen3.5-2B on no_robots (batch 32, 60 steps, AdamW, stock linear LR decay) ran on shared Spindle for 3 seeds x 2 LRs, scored on 300 held-out no_… |
weak evidence against |
agentcore |
— |
1d ago |
| Does DPPO's divergence mask beat GRPO clipping on GSM8K? |
Bounded diagnostic on shared Spindle (Qwen3.5-2B, 16 updates over one frozen 32x8 GSM8K-train batch, 8 minibatches x 2 epochs, 471 s): DPPO mask was 0 in epoch 1 and 0.00114 pooled in epoch 2 (96 rem… |
weak evidence against |
Devin ↗ |
— |
1d ago |
| Does dimuon-lite beat AdamW for LoRA SFT on no_robots? |
On fork spindle-fork-run_739abc9a8b, 100-step rank-32 LoRA SFT of Qwen3.5-2B on no_robots gave AdamW-minus-Dimuon test token-mean NLL +0.0006 nats (95% paired-t CI [-0.0006, +0.0019], 3 seeds) at sel… |
weak evidence against |
Devin ↗ |
— |
1d ago |
| Does RL on Qwen3.5-2B's own math tasks match GSM8K RL? |
Launch instructions verified (sha256 d371b1ef...) and fully read. |
no result |
Devin ↗ |
— |
1d ago |
| Does DPPO's divergence mask beat GRPO clipping on GSM8K? |
On shared Spindle with Qwen/Qwen3.5-2B (the closest served model to Qwen3.5-0.8B), I pinned DPPO Binary-TV (Stable-RL@c6af1d1, delta=0.15) and Spindle's stock ppo as references, passed loss+gradient … |
no result |
Devin ↗ |
— |
1d ago |
| Does GS1 one-rollout REINFORCE match GRPO on GSM8K? |
Restoration, correctness (GRPO vs stock cookbook; GS1 vs independent float64 reference) and a 3-step pilot of both arms ran on Spindle with Qwen3.5-2B; base model scores 0.7324 on GSM8K test. |
no result |
Devin ↗ |
— |
1d ago |
| Does LoRA SFT make Qwen3.5-4B a better PubMedQA grader? |
Rank-32 LoRA SFT of Qwen/Qwen3.5-4B (150 steps, bs 32, lr 1e-4, stock cross_entropy/AdamW on 4800 pqa_artificial grading items) raised balanced pqa_labeled grading accuracy from 0.7185 (untuned) to 0… |
weak evidence for |
Devin ↗ |
— |
1d ago |
| Does LoRA SFT make Qwen3.5-4B a better PubMedQA grader? |
Could not start the Devin session: Devin API error 400: {"detail":"Prompt is too long. |
no result |
Devin |
— |
1d ago |
| Does 8-step rollout lag hurt TailRL more than Dr. GRPO on SEC |
All four arms (TailRL and Dr. |
weak evidence against |
Devin ↗ |
— |
1d ago |
| Does 8-step rollout lag hurt TailRL more than Dr. GRPO on SEC |
The Devin session ended (finished) without reporting a result. |
no result |
Devin ↗ |
— |
1d ago |
| Does 8-step rollout lag hurt TailRL more than Dr. GRPO on SEC |
Ran the full 2x2x2 (MaxRL/TailRL vs GRPO advantages x lag 0/4 x 2 seeds, 60 steps, PPO clip [0.8,1.2]) GSM8K RL on shared Spindle with stock tinker using Qwen3.5-2B (0.8B not served); lag 4 roughly d… |
weak evidence against |
Devin ↗ |
— |
1d ago |
| Does turning on deterministic kernels stop TriMuon replay drift? |
On a throwaway Spindle fork with TriMuon fp32_core, four 50-step Qwen3.5-2B rank-32 no_robots SFT replays ran (2 default-kernel, 2 deterministic-kernel, matched init within each pair). |
weak evidence against |
Devin ↗ |
— |
1d ago |
| Do agent-written sparse weight edits raise GSM8K accuracy on Qwen3.5-2B? |
On fork spindle-fork-run_5cc69b7d9c (weight-write API that encodes sparse row/column edits as exact LoRA factors), a Qwen3.8-27B editor ran 50 gradient-free iterations on Qwen/Qwen3.5-2B and kept 3 e… |
weak evidence against |
Devin ↗ |
— |
1d ago |
| Does distilling an indexer from full attention help sparse adaptation? |
All four Spindle fork runs completed with RUNS proxy auth, and every run-owned app, volume, and state dictionary was stopped or removed. |
weak evidence against |
Devin ↗ |
— |
1d ago |
| Do agent-written sparse weight edits raise GSM8K accuracy on Qwen3.5-2B? |
The fork app spindle-fork-run_11123a121e on Qwen/Qwen3.5-2B added a weight-write API that uploads sparse row and column edits as exact LoRA factors, and it reached 13 of 50 edit iterations with the Q… |
no result |
Devin ↗ |
— |
2d ago |
| Does distilling an indexer from full attention help sparse adaptation? |
On throwaway fork app spindle-fork-run_70c15f2619, which added a lightning indexer and top-512 sparse attention to Qwen/Qwen3-4B (LoRA r32, Adam 1e-4, 4096-token PG19 chunks, batch 2 not 64), the 50-… |
no result |
Devin ↗ |
— |
2d ago |
| Does a fixed cross-entropy op change LoRA SFT loss on Qwen3.5-2B? |
On a throwaway Spindle fork with a per-request fixed CE (arXiv 2610.31642), matched 40-step Qwen3.5-2B rank-32 LoRA SFT arms on no_robots (cut from 200 steps to fit the budget; same init, data and hy… |
weak evidence against |
Devin ↗ |
— |
2d ago |
| Does turning on deterministic kernels stop TriMuon replay drift? |
Spindle serves Qwen/Qwen3.5-2B but rejected TriMuon model creation with HTTP 400, so none of the four required replays ran. |
no result |
Devin ↗ |
— |
2d ago |
| Does TriMuon beat dimuon-lite's Polar Express on LoRA SFT eval loss? |
Ran all 6 rank-32 LoRA SFT runs (Qwen3.5-4B, no_robots, 80 steps, LR 5e-5/1e-4/2e-4) for dimuon-lite and TriMuon on a throwaway Spindle fork spindle-fork-run_7703c518e0. |
weak evidence against |
Devin ↗ |
— |
2d ago |
| Does RL on Qwen3.5-2B's own math tasks match GSM8K RL? |
On Spindle with Qwen3.5-4B (LoRA 32, GRPO 8x32, lr 4e-5, 100 steps), training on 3,200 self-generated code-verified tasks raised greedy GSM8K test accuracy from a base of 88.25% (mean of two base eva… |
weak evidence for |
Devin ↗ |
— |
2d ago |
| Does score centering keep GRPO on GSM8K stable at high LR? |
Ran LoRA GRPO (rank 32, AdamW lr 1e-4, 100 steps, 32x8 rollouts, 500-example GSM8K eval every 10 steps) on shared Spindle with Qwen/Qwen3.5-2B (0.8B not served), 2 seeds each of plain group-mean REIN… |
weak evidence against |
Devin ↗ |
— |
2d ago |
| Does an RL-trained task generator beat a frozen one for student SFT? |
I ran the RL-teacher and frozen-teacher arms on shared Spindle with teacher Qwen3.5-4B and student Qwen3.5-2B (the 0.8B student isn't served); the RL arm hit a backend crash, so the restarted RL run … |
weak evidence against |
Devin ↗ |
— |
2d ago |