A public, auto-synced log of every verified proof-of-training run merged into SparkDistill — the SN74/Gittensor miner economy for Triton-kernel-specialized LLM distillation.
This mirrors runs/ledger.jsonl,
SparkDistill's own immutable, append-only record — that file is the source of truth,
written by SparkDistill's CI (eval.record_training_ledger) at merge time; this repo
only renders a public-facing copy of it (render_log.py, synced on a schedule by
.github/workflows/sync.yml). Nothing here is written by
hand, and nothing here can drift ahead of the ledger — only behind it, until the next sync.
Label: eval:BASELINE (2×, first verified run on a GPU architecture — nothing to
compare against yet), eval:XS–eval:XL (verified quality tier vs. the current
frontier), or eval:REJECT (did not pass verification). See SparkDistill's
miner guide
for what these mean and how they're computed.
| Date | Run ID | PR | Label | Best benchmark | Δ vs frontier | Attested | Regressions | Proof |
|---|---|---|---|---|---|---|---|---|
| 2026-07-29 | 2026-07-29-philluiz-blackwell-xl-v1 | #318 | eval:XL | triton | +21.2% | yes | — | proof |
| 2026-07-29 | 2026-07-28-qwen3.5-4b-blackwell-seq8192-v1 | #301 | eval:XL | triton | +41.8% | yes | — | proof |
| 2026-07-28 | 2026-07-28-magicrails-blackwell-v3-think | #288 | eval:REJECT | — | — | yes | — | proof |
| 2026-07-15 | 2026-07-15-magicrails-hopper-v2 | #120 | eval:BASELINE | — | — | yes | — | proof |
| 2026-07-11 | 2026-07-11-qwen3.5-4b-mining-001 | #40 | eval:BASELINE | gsm8k | — | yes | — | proof |
Moving fast is key to success.