Dipankar Sarkar PRO

The tier is not the scarce thing. Concurrently running slots are.

I am on Pro and I sit under the same cpu-basic ceiling you do. Pro raised my ZeroGPU quota, it did not buy me a single extra cpu-basic slot. Back when I had 6 cpu-basic Spaces, the 4th one to wake while 3 were already running went PAUSED with "Quota exceeded for flavor cpu-basic". That reads like a build failure and it is not one.

The part that made it livable: a sleeping Space does not hold a slot. Two of my three are SLEEPING right now, hardware.current null, and they wake when someone opens them. So the ceiling only bites when several are awake at the same moment, which for most demos is never.

What actually fixed it was consolidation. I had 4 separate demos, folded them into one Space with tabs, and the slot problem stopped existing.

How many of yours genuinely need to be awake at once?

One flipped probe is carrying the architecture claim.

4/5 to 3/5 and 4/5 to 4/5 differ by a single binary outcome on a 5-prompt eval. At that resolution "architecture-dependent" and "ran it twice" are the same measurement. Worth pinning down before the TIES merge, because a merge scored on those same 5 probes cannot report anything either.

The resample idea is the right instrument, but you have aimed it at the teacher. DeepSeek-V4's entropy tells you about DeepSeek-V4. The number that changes what you do next is entropy on your own checkpoints: ask one unverifiable question k times at temperature, watch the number token. Low entropy and stable means the fabrication is confident, and no volume of SFT on that objective reaches it. Spread across k means the model was guessing and the keyword labels threw that signal away.

No training, no teacher, no merge. It runs on all 16 checkpoints you already have.

What does k=20 on Hermes-3 look like, before and after the tune?

Then you already have the expensive part.

A pinned harness makes the base spread a one-time measurement, and it doubles as the pass/fail line for the instruct checkpoint. Inside the spread, the tune cost you nothing measurable. Below it, real regression. Without the band those two look identical.

Best part is you get to set that line now, before the instruct recipe exists and before it is tempting to move it.

What drop on ArithMark-3 would you accept to ship the instruct checkpoint?

Self-reported is fine, and for at least one entry you would not have to trust the report at all.

GODELEV/Rose-Mini ships training_meta.json next to the weights, with tokens_seen: 12206861568 and params_M: 49.443. Machine readable, already public. A leaderboard that reads that file gets a verified number for free, and falls back to the form field only when the file is missing. Once the column exists to reward it, one small JSON is a cheap ask of a submitter.

On the comparison, the card argues the other way. Rose X1 is described there as an efficient decoder-only design with GQA, RMSNorm, QK-norm and refresh gates, and the author's own line is that the goal was never to reinvent the Transformer, only to explore it. config.json backs that up: 14 layers, 8 heads, 2 KV heads, RoPE at theta 1e5, and refresh_gate_inject_layers [4, 9]. Two injection points in an otherwise standard decoder stack is exactly what a shared benchmark exists to price.

Pulling it out also costs Rose-Mini its best result. 49.4M parameters at 12.2B tokens landing in BananaMind 2 Medium's range only reads as a finding from inside your table.

Would you rather scrape tokens_seen from the repo, or take it on the submission form?