s6-micro-v1

Failed its pre-registered gate: published as a result, not for use. The gate (docs/s6-gen-prereg.md Β§6) asked for at least 0.45 on typed_decisions_bench, an ECE of at most 0.08 over the test sources and no game more than 5% below s3-pico. s6-micro-v1 reached 0.547 on the bench and plays every game at s3-pico's score, but its ECE is 0.0821 at full depth (0.0804 at its conformal thresholds) against the guard of 0.08: it failed its calibration guard. The file is published so that the result can be checked and run (owner's decision of 15/11/2026, amendment 15 of the pre-registration); the verdict does not change. Numbers: BENCH.md, section skycmd-s6.

A 2.4 MB ternary model for four games and for typed questions in free text: the larger sibling of s6-nano. 5,587,201 parameters. It reads a state as text and answers typed questions (choice, score, noul) with probabilities, one forward pass per question. It is published as s6-micro-v1. The second version pre-registered in amendment 12 (360 minutes of decision training instead of 180, the same gate) ran on 15 and 16/11/2026 and passed, on a second look at the same test records (amendment 17); it is published as s6-micro. This file's verdict does not change.

Results

Test sets, at most 2,000 questions per source, ternary GGUF of the last DAgger round, full depth (Ο„ = 1); the reference rows are the same records:

set s6-micro-v1 s6-nano (passed) reference
typed_decisions_bench 0.547 0.480 majority class 0.398; s3-pico 0.382, s4-micro 0.352, s5-lite 0.696
MASSIVE intents, English 0.518 0.425 s3-pico 0.187 (see the limitations)
MASSIVE intents, French 0.275 0.239 s3-pico 0.117
When2Call (test split of a trained source) 0.430 0.404 s3-pico 0.368
NOTAM subject / NOTAM multi (never trained on) 0.052 / 0.394 0.055 / 0.493 s3-pico 0.163 / 0.319
all sources 0.378 0.360
ECE over all these sources 0.0821 0.0479 the gate allows 0.08
  • At its conformal thresholds (not published): 0.544 on the bench, ECE 0.0804, 4.59 block passes on average instead of 6, early exit on 36.0% of the questions. Still above 0.08.
  • Where the ECE comes from (read after the gate, exploratory, amendment 12): the largest part is notam:notam_multi (ECE 0.2558, against 0.0710 for s6-nano). On one NOTAM statement, false for 99.5% of its 412 items, s6-micro-v1 answers "true" for almost all of them. The model is under-confident on the bench (mean confidence 0.4815 against an accuracy of 0.5470) and over-confident on When2Call (0.5372 against 0.4305). The 95% bootstrap interval of the pooled ECE is [0.0746, 0.0900]; the gate is read on the point value, as pre-registered.
  • ECE by source: bench 0.0660, MASSIVE en 0.1334, MASSIVE fr 0.0313, When2Call 0.1176, NOTAM subject 0.0636, NOTAM multi 0.2558.

Closed loop (sky-arena, 30 episodes per game, the same seeds as s3-pico, full depth): Flappy 200.00, Rescue 3.00, Dodge 9.47, Doom-like 3.27 mean score, the scores of s3-pico and s6-nano.

Latency: median 7,346 Β΅s per decision natively over the test sources at full depth (s6-nano: 1,488 Β΅s). The WASM line of the run's results file was measured with early exits against a full-depth native dump (amendment 13), so it does not describe this file and is not quoted; the WASM latency of this file at full depth was not measured.

Full depth

The file answers at full depth: all 6 block passes, every time. Its three exit thresholds are set to 1 in the GGUF (skycmd-s2.exit_threshold.{choice,score,noul}), as amendment 13 of the pre-registration defines a publication at full depth; the weights are those of the evaluated file (the last DAgger round), byte for byte. You can still pass another threshold at run time (setExitThreshold in the browser).

What it does well, and badly

  • Better than s6-nano on accuracy: 0.547 against 0.480 on the bench, the announced goal of 0.50 met, and higher on MASSIVE and When2Call.
  • Worse calibrated: 0.0821 against 0.0479, and the guard is what the gate checks. Its probabilities are not reliable across sources: do not read them as confidences.
  • Slower and larger: a median of 7,346 Β΅s per decision natively against 1,488 Β΅s for s6-nano, in a file of 2,420,032 bytes against 715,136.
  • It cannot read NOTAMs (0.052 on the subject question), and in the games it scores what s3-pico and s6-nano score, no more.

Architecture

The s4-micro shape (skycmd-s2 GGUF tag): question and options form a static stream encoded once and cached; the state goes through weight-shared blocks with exits after 2, 4 and 6 block passes; options never attend to each other; ternary weights trained with QAT. d 256, 4 heads, vocabulary 4,096, windows of 384 state tokens and 128 question tokens. The pre-registered ablation adopted none of the s6 additions (Parallax, Muon v2, XSA and others), so this model uses softmax attention and AdamW.

Limitations

  • Failed its gate. Recommended for no use. s6-nano, which passed the same gate, and the second version s6-micro, which passed it on a second look, are the models to try.
  • The English MASSIVE score is not strictly zero-shot: 206 of the 2,000 test utterances also appear in the training split, 201 of them in a source derived from MASSIVE (0.495 on those 206, 0.521 on the 1,794 others). None of the 1,000 French ones does.
  • One run, one seed. The decision training was stopped by hand at step 27,280 after about 34 minutes in swap and resumed from step 27,161 uncompiled (amendment 11); the steps done in swap are kept, so it has fewer steps than its 180 minutes would otherwise give.
  • The games are civil drone scenarios and a Doom-like shooter drawn with Freedoom art; the model only advises and coded rules always keep a veto. This is a demo and a tutorial, not flight software.

Usage

Runs in the skycmd engine only (browser, Node, Python, CLI, server):

cargo install --git https://github.com/maxgfr/skycmd sky-serve
skycmd decide --model s6-micro-v1.gguf --request req.json --pretty

Request format and bindings: install guide.

Training

From scratch in Rust (mlx-rs) on an Apple M5, in a braked mode, so the budgets are working minutes: masked-language pretraining in float (180 minutes, 31,700 steps, 0.52B tokens, 93 per parameter), then decision training with ternary QAT after a 10 % float warm-up (180 minutes, 38,826 steps, learning rate Γ—0.5 chosen by a pre-registered pilot) and three DAgger rounds on the games (20 minutes each: 5,951, 5,898 and 5,952 steps). Data: the typed-decision sets of Hugging Face, classic datasets converted into typed questions, game states labelled by exact oracles and a numeric curriculum, the same as s6-nano. Pretraining text: SimpleStories, TinyStories, FineWeb and the non-web parts of Dolma 3 and TxT360. Soft labels on 747,077 training questions come from GLM-5.3-flash (open weights, MIT, called through the z.ai API), kept only where they agree with the gold answer; no test question was sent. The teacher's weights are open (MIT); the labels and their publication are described in DATA_LICENSES.md, with every other license.

Files

  • s6-micro-v1.gguf: ternary weights, tokenizer, temperatures, and exit thresholds set to 1 (full depth). 2,420,032 bytes, sha256 33b160fc6491d0f9d198e1e0f6eaed2f7a7c2d82d69fdce384ad9b60814913c7.
Downloads last month
49
GGUF
Model size
2.21M params
Architecture
skycmd-s2
Hardware compatibility
Log In to add your hardware

We're not able to determine the quantization variants.

Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Datasets used to train maxgfr/s6-micro-v1

Spaces using maxgfr/s6-micro-v1 2

Collection including maxgfr/s6-micro-v1