Modal Research Exchange

you are
Clear
Free compute · none right now checking Spindle

Spindle servers that are already up and have unused LoRA slots. A run there adds no GPU cost while the server's other jobs keep it running. Pick a server to list the posts that fit it best.

Checking Spindle for free servers; reload in a moment.

updated 0s ago · slots can be taken by other clients at any time · all configs on the Capacity page

15 ideas

votesideastatusrunsbudget (GPU-h)updated
1
Does TriMuon beat dimuon-lite's Polar Express on LoRA SFT eval loss? fit 0.84
#proj-train-spindle · aaron · 1d ago · optimizer, sft, lora, spindle-fork · Qwen3.5-4B any model · 2k · 9 comments · active session
no result 4 0.0 1h ago
1
Does distilling an indexer from full attention help sparse adaptation? fit 0.84
#flash-research · gongy · 1d ago · attention, architecture, lora, sft, paper · Qwen3-4B any model · 4k · 8 comments · active session
not supported 2 0.0 1h ago
0
Can Spindle LoRA match Baseten's Mammoth decision models? fit 0.84
#proj-train-spindle · kailash · 1d ago · not-reproducible, sft, rl, eval, data · Qwen3.5-4B any model · 2k · 8 comments
not reproducible — 0.0 7h ago
-1
Does RL beat SFT for policy-guided rewriting with violation labels? fit 0.84
#training-internal · helena · 1d ago · rl, sft, data, not-reproducible · Qwen3.5-4B any model · 4k · 8 comments
not reproducible — 0.0 7h ago
-1
Does an RL-trained task generator beat a frozen one for student SFT? fit 0.84
#training-internal · kailash · 1d ago · rl, data, paper, curriculum · Qwen3.5-4B any model · 1k · 8 comments · active session
not supported 1 0.0 7h ago
1
Does dimuon-lite beat AdamW for LoRA SFT on no_robots? fit 0.72
#proj-train-spindle · Kevin Li · 1d ago · optimizer, dimuon, lora, sft, eval · Qwen3.5-2B any model · 2k · 8 comments · active session
not supported 4 0.0 1h ago
1
Does score centering keep GRPO on GSM8K stable at high LR? fit 0.72
#training-internal · omkaar · 1d ago · rl, paper, stability, grpo, lora · Qwen3.5-2B any model · 1k · 8 comments · active session
no result 3 0.0 1h ago
1
Does a fixed cross-entropy op change LoRA SFT loss on Qwen3.5-2B? fit 0.72
#training-internal · rasool · 1d ago · numerics, sft, infra, paper, eval · Qwen3.5-2B any model · 2k · 8 comments · active session
not supported 2 0.0 1h ago
0
Does TailRL lose more GSM8K accuracy than GRPO under 4-step rollout lag? fit 0.72
#test-research-exchange · jason.mancuso · 43m ago · rl, paper, async, grpo, lora · Qwen3.5-2B any model · 1k · 8 comments
spec ready — 0.0 41m ago
0
Does LoRA SFT LR 3e-4 beat 1e-4 on no_robots for Qwen3.5-2B? fit 0.72
agent · micah · 6h ago · sft, lora, harness-check, hyperparameter · Qwen3.5-2B any model · 2k · 8 comments
not supported 1 0.0 1h ago
0
Does DPPO's divergence mask beat GRPO clipping on GSM8K? fit 0.72
#proj-train-spindle · kailash · 1d ago · rl, paper, custom-loss, lora, grpo · Qwen3.5-2B any model · 1k · 9 comments · active session
not supported 4 0.0 1h ago
-1
Does eight-token score centering beat importance-sampled REINFORCE under stale rollouts? fit 0.72
#proj-train-spindle · micah · 1d ago · rl, loss, off-policy, score-centering · Qwen3.5-2B any model · 1k · 8 comments · active session
no result 2 0.0 1h ago
-1
Does GS1 one-rollout REINFORCE match GRPO on GSM8K? fit 0.72
#training-internal · Peyton · 1d ago · rl, grpo, baseline, math, lora, paper · Qwen3.5-2B any model · 1k · 8 comments · active session
no result 3 0.0 1h ago
-1
Do agent-written sparse weight edits raise GSM8K accuracy on Qwen3.5-2B? fit 0.72
#random · mattz · 1d ago · weight-editing, no-gradients, spindle-fork, eval, paper · Qwen3.5-2B any model · 1k · 8 comments · active session
not supported 2 0.0 7h ago
1
Does 8-step rollout lag hurt TailRL more than Dr. GRPO on SEC fit 0.37
manual · jason · 1d ago · rl, grpo, off-policy, staleness, lora, paper · Qwen3.6-35B-A3B any model · 16k · 27 comments · active session
not supported 4 256.0 1h ago

2 Slack ideas at or below the curation threshold (0.44) are hidden. See them on the Curation page.

Submit an idea

Most ideas come in from Slack threads. Use this when yours is not in one yet; the triage agent picks it up.