Skip to content

Latest commit

 

History

History
195 lines (151 loc) · 10.6 KB

File metadata and controls

195 lines (151 loc) · 10.6 KB

Scoring and Rewards

PRISM is a research lab. The emission path scores one thing: which package produces a better model for us after fair from-scratch learning — measured first by held-out / generalization, then by recomputed prequential bits-per-byte. The challenge owns every number.

Surface honesty (locked).

Surface Role Ranking
Emission crown (this page: leaderboard → raw weights) Subnet reward eligibility Held-out / generalization PRIMARY, prequential bpb SECONDARY (Official-like invert)
Official Comparison / multimetric.v1.1 / Complete View Published scientific miner architecture grade Multi-axis vector + polar honesty; not the emission scalar in v1

Do not claim multimetric / Complete View silently replaces emission. Prior LAB-GPU K=1 short-ctx wins remain provisional only. Challenge-owned prism_train_series.v1 is visibility + residual densify for sample-eff/stability only — never sole primary over held-out/bpb. When an Official grade pin requires the series and it is missing/corrupt, Official grade fail-closes (not silent PASS); miner dashboards remain non-authoritative. Worker-plane integrity uses constation (constation_ok over the six-mechanism bundle). REAL-PROVIDER TEE is retired for Prism product (historical tables may still say BLOCKED; never a production scoring gate). Under P1, a missing or failed constation bundle means no score row at all, so attestation is decisive for score presence (not a soft label orthogonal to ranking).

flowchart LR
    Loss[Single-pass online loss stream] --> Bpb[Prequential bits-per-byte]
    Loss --> Hold[Held-out generalization]
    Hold --> Final[emission_rank / final_score]
    Bpb --> Final
    Gap[Train-vs-held-out gap - memorization penalty] --> Final
    Anomaly[Step-0 smuggled-weights anomaly multiplier] --> Final
    Final --> Board[Leaderboard ordered by emission rank DESC]
Loading

Emission Primary: Held-Out / Generalization

For the production leaderboard and raw-weight path, the ranking primary is held-out / generalization quality the challenge computes itself on the secret val split after forced-init re-execution.

Preferred primary form (higher better):

heldout_delta = bpb(random-init twin on val) - bpb(trained model on val)

Alternate pin option: absolute val_bpb_trained (lower better) when twin delta is unavailable. Honestly lower held-out free energy (or larger honest improvement over the random-init twin) ranks higher. A train-only memorizer must not win the emission crown on primary.

With multi-seed residual when configured:

primary_S = mean_k heldout_delta(S; seed_k)     # higher better

(or the documented alternate mean of val_bpb_trained, lower better).

final_score / emission rank folds held-out primary first, then secondary bpb, then anti-memorization and step-0 multipliers, so the leaderboard's descending order ranks better generalizing learners first.

Emission scoring is architecture-agnostic: the same fold runs for every admitted nn.Module that clears AST + dual param ladder. There are no family-specific emission shortcuts (Transformer / pure-torch SSM / looped-depth / novel hybrids receive the same held-out primary + bpb secondary path). Lab seeds tiny-1m and mamba-tiny are default exploration shapes only — not privileged emission families.

Emission Secondary: Prequential Bits-Per-Byte

During the re-execution, the challenge feeds the model fresh, single-pass batches from the locked train split and records its loss on each new batch before the optimizer updates on it. Single-pass data makes this online (predict-then-train) loss the prequential code-length by construction. The challenge integrates that code-length and normalizes it by the raw UTF-8 bytes covered:

bpb = (sum over consumed tokens of -log2 p(token)) / total_bytes_covered

Byte normalization makes the metric tokenizer-agnostic (any tokenizer compares like for like); integrating the whole curve defeats single-checkpoint gaming; scoring each token before training on it removes held-out leakage by construction; and forced random init makes smuggled pretrained weights inert.

Lower bits-per-byte is better as the secondary emission axis. When held-out is within a small epsilon (near-tie on primary), prequential bpb breaks the tie so a strictly better recomputed compression signal can reorder near neighbors — but a clearly better held-out winner is not overturned by bpb alone.

Continuity transform (still used where products report a scalar display fold of bpb before penalty):

bpb_display = 1 / (1 + bpb)        # higher display = lower bpb; secondary only under emission rank

Compute Normalization, Not Wall-Clock

The score is compute-normalized: normalized by tokens consumed (and optionally estimated FLOPs), never by wall-clock time. A faster GPU or more GPUs cannot buy a better score; wall-clock is only a safety cap. This keeps scores fair across the 1-to-8 GPU range even though the scored run uses one physical GPU.

Dual param ladder (explore → promote)

Emission crowns respect the GPU-limited small-first ladder:

Stage Cap Crown role
Explore / provisional 124M Qualifying scores may provisional-crown architecture or training pools
Promote / final 350M Same package/family pin re-eval confirms or revokes the provisional crown

Explore is the default thrash size. Durable supremacy requires promote confirmation; a failed or losing promote revokes the provisional crown so dead provisional winners do not keep emission weight.

Anti-Memorization Gap

The challenge measures the train-vs-held-out gap (converged train bpb vs held-out val bpb on the same byte basis). An excessive gap flags memorization and multiplies a penalty into the emission score, so a memorizer ranks below an equivalent non-memorizing learner. The comparison is basis-consistent so a benign learner is not falsely flagged.

Anomaly Zeroing

A step-0 / smuggled-weights anomaly (an impossibly low initial loss under forced random init) drives the anti-cheat multiplier to zero, so an anomalously good looking compression curve is zeroed rather than rewarded. A degenerate run (zero coverage, non-finite, or out-of-band bpb) is failed rather than scored.

Leaderboard And Weights

The emission leaderboard ranks by held-out-primary emission rank (with bpb secondary and folded penalties). Remaining ties break by earliest-commit-wins, then submission id, for a total, reproducible order. Each hotkey appears at most once, keeping its best submission.

PRISM converts completed scores into raw hotkey weights and pushes them to the BASE master for aggregation. Two-tier ownership defaults are architecture 0.50 / training 0.50 (no silent 0.60/0.40 or 0.65/0.35 drift on defaults). Both tiers consume the same emission rank metric. get_weights remains available for inventory/compatibility. On-chain set_weights is validator-owned only; PRISM never writes weights on-chain.

Scientific surfaces (not emission scalar)

Offline Official Comparison is the scientific miner architecture grade: multi-axis Official / Complete View with held-out primary, bpb secondary, long-ctx, reasoning, and polar honesty (TIE_POLAR / crown_allowed=false when axes disagree). The additive multimetric scorecard annex (scorecard_id=multimetric.v1.1) and Complete View publish the scientific vector. They remain published research grade and do not silently replace this emission crown path in volume-1 (a future phase may fold axes into emission only with an explicit product bump).

Challenge-owned prism_train_series.v1 time-flow (loss/bpb + mandatory grad_norm / clip when instrumented) densifies sample-eff/stability residual and operator visibility. Series must never sole-rank mining packages over Official held-out or recomputed bpb axes, and must not substitute emission held-out primary.

Source Of Truth

Every number above is recomputed by the challenge from the challenge-authored prism_run_manifest.v2.json. Miner-reported metrics and miner-written manifests are ignored. The legacy raw-loss term and the v1-NAS architecture/training ownership pools are retired from the score.

Miner self-reports remain non-authoritative on both the emission path and Official Comparison mode (including scorecard v1.1, Complete View, and train series).

Constation and score presence (P1 fail-closed)

Worker-plane Lium runs are admitted to the emission path only when Prism accepts a full constation bundle and constation_ok is true. That predicate is the sole route to effective_tier == 1 (ceiling 1; no TEE). Self-reported image digests and historical "IMAGE_PIN match" labels do not elevate tier alone and do not create a score row when the bundle is missing or fails.

Outcome Score row? Ranking impact
Valid bundle, constation_ok Yes (then held-out/bpb ranking applies) Eligible for emission crown
Missing bundle No (miner_fault:missing_constation_bundle) Not ranked
Failed mechanism(s) No (miner_fault:<code>) Not ranked
Infra outage after retries No (infra_fault:*) unless audited break-glass Break-glass admits at tier 0 only

So constation is not orthogonal to ranking: without a score row there is nothing to rank. Among admitted rows, ranking remains held-out primary / bpb secondary as above. Trust model detail and the six mechanisms with honest limitations: Security and prism-recipe security. Official Comparison notes: Official Comparison §17 train series telemetry and scorecard honesty.

Reference Studies

  • Prequential / online coding — Dawid, 1984: score the integrated predict-then-train loss, not a final checkpoint.
  • Minimum description length — Rissanen, 1978: treat compression (code-length) as the learning signal.
  • Scaling laws — Kaplan et al., 2020: compare loss trajectories under matched compute.
  • Compute-optimal scaling — Hoffmann et al., 2022: normalize by tokens/compute so under- or over-trained regimes do not skew ranking.
  • Dataset provenance — Penedo et al., 2024 (The FineWeb Datasets): freeze the data revision and shards for reproducible official runs.
  • Generalization over train compression — held-out primary emission prefers architectures that actually improve secret val free energy under matched train budgets.