a bright data hackathon submission · 2026

Forge a
training recipe
that actually runs.

A live compatibility graph for ML training components. Pick your optimizer, scheduler, quantization, technique. Incompatible tools grey out instantly with cited evidence from real practitioner runs. Greyed means not yet satisfied, not forbidden — add the missing ingredient and watch it re-light.

T1 QLoRA T1 Full FT T1 bnb 4-bit needs · add T1 LoRA T1 Llama-3 T2 Grad Ckpt
what it does

A loadout-builder for ML training.

Pick optimizers, schedulers, quantization, techniques. The graph re-resolves live and tells you what fits, what's missing, and what blows up — every fact backed by a citation.

I

Compatibility-aware

Pick a component and every other one re-styles by status: available, conditional, or blocked. Greyed means not yet satisfied — add the missing ingredient and it re-lights in real time.

II

Evidence-backed

Every grey-out, every fix, every conflict carries its cited source — verbatim quote, URL, and tier badge. Hover anything to see where the claim comes from.

III

Benchmark-backed

Scheduler picks aren't picked from a paper — they're picked from real measured val_loss / val_acc on distilbert / sst2, three seeds each. DLRS leads. OneCycle is flagged above cohort cutoff.

how to use it

Three strikes on the anvil.

Sixty seconds, three motions. The demo path the team built it around.

Step One

Pick from the inventory.

The left rail is your loadout. Tap any optimizer, scheduler, technique, quantization, architecture, or inference engine. The instant you equip one, the graph re-runs POST /resolve and re-styles every other node by status.

"Pick QLoRA. Full Fine-Tune slams to blocked. bnb 4-bit shimmers conditional — ‘needed by QLoRA’."
QLoRA
equipped
T1
Full Fine-Tune
blocked · breaks with QLoRA
T1
bnb 4-bit
needed by QLoRA
T1
Step Two

Read the recipe.

The right rail is your crafting window. Conflicts are scored in rust. Unmet requirements glow amber with a one-tap + add bnb 4-bit button. Benchmark-backed scheduler picks list themselves by measured val_loss with the leader marked with a star.

Every entry shows the tier badge and a source link — so the credibility chain stays visible all the way through.
QLoRA Full Fine-Tune
QLoRA freezes the base model; Full Fine-Tune updates it. Pick one.
T1 arxiv.org/abs/2305.14314
QLoRA needs bnb 4-bit
QLoRA is defined as 4-bit NF4 base + LoRA adapters — load the base in bnb 4-bit.
Step Three

Inspect the evidence.

Hover any node, edge, or recipe entry to surface the verbatim quote from its source and the tier badge it earned. T1 is Rick's own verified runs (immutable). T2 is community-corroborated. T3 is single-source — lives in the review queue until something else agrees.

The point of the tooltip is the source URL. The point of the graph is the decision. The point of the tier is so you know how much to trust both.
AdamW 8-bit blocked
breaks with Per-Layer LR Rotation
Pass custom optimizer_grouped_parameters — adamw_8bit's default decay/no-decay split silently no-ops per-layer LR.
"adamw_8bit produces decay/no-decay param groups by default; per-layer LR rotation silently no-ops unless you pass them yourself."
mempalace · rick T1
trust model

Every claim earns its tier.

We don't assert nonsense as fact. Three tiers, three colors, three weights. The scraper feeds the bottom; corroboration promotes upward.

T1
Legendary · 1.0

Verified runs

Rick's own research notes and reproduced runs. Immutable — the scraper can never overwrite these, only add corroboration.

e.g. Per-Layer LR Rotation × AdamW 8-bit · mempalace://chaos-injection-trainer-notes
T2
Rare · 0.8

Community-confirmed

Official docs, or two-plus independent sources agree. The bulk of the workable graph.

e.g. Muon → AdamW 8-bit · kellerjordan.github.io/posts/muon
T3
Common · 0.5

Single-source

One scraped report, no corroboration yet. Lives in the review queue. Promotes when something else agrees, or when Rick signs off.

e.g. Lion × bnb 8-bit divergence · github.com/bitsandbytes issue
the live data layer

The graph grows itself.

A Bright Data scraper hunts SERP results in English and Chinese, extracts (A, relation, B, conditions, fix, quote) tuples with Haiku, resolves them against canonical node aliases, and proposes them to the review queue.

Watch the Live Discoveries hearth at the bottom of the Forge while you work — new edges literally rise out of it as the scraper corroborates them. The graph is alive.

English-only sources miss large chunks of the Chinese ML ecosystem. We cover both. The scraper writes nothing to T1 — that's Rick's territory. Single-source claims sit at T3 in the queue and only ascend with evidence.

SERP·en/zh arxiv · github
scrape_as_markdown bright data
extract tuple haiku
entity-resolve canonical aliases
review_queue T3
corroborate T2

Light the forge.

Ninety seconds. Ten components. One recipe that actually compiles.

Enter the Forge