AI Hackathon
← DEPARTURES
AI SafetyStatus: Ended

Lexsi AI Safety and Alignment Hackathon

Lexsi Labs, Cohere Labs, Hugging Face · on Lexsi

Cash prize: Not stated
Free credits: Credits for winners
Who can enter: Solo OK Country rules checked
Code rights: L2 Promotion
Fine-tuningRed teamingAgentic fine-tuningCohereHugging Face
Submissions closedNov 16

This hackathon has ended. Submissions closed on 2026-11-16 11:59 UTC.

Red-team, fine-tune, and audit open-weight language models in a fully online, 14-day AI safety and alignment build.

We checked the deadline, prize and entry rules against the organizer's page. The summary, ideas and FAQ were written by AI from those facts.

One email a week

New events that match your topics, every Monday. Unsubscribe anytime.

Flight details

Registration closed:
2026-11-09 00:00 UTC
Submissions closed:
2026-11-16 11:59 UTC
Cash prize:
Not stated
Prize details:
No cash prizes. The top five places get $500 to $1,750 each in Lexsi Labs and Cohere credits, plus six months of Hugging Face Pro for every teammate.
Free credits:
Lexsi Labs and Cohere credits, Hugging Face Pro for winners.
Who can enter:
Open to builders worldwide. Fully online, Nov 1 to Nov 22.
Code rights:
Teams keep ownership of what they build. The organizers may showcase submitted projects for event and promotional purposes.

Prizes listed: $30,000+ in prizes and credits combined. The cash part is not stated. First place is worth $3,662 in credits plus six months of Hugging Face Pro per teammate.

Before you board

  • Registration closed on Nov 8 (extended by 12 hours). Only registered teams can still submit, by Nov 15 at 11:59 PM Anywhere on Earth.
  • Teams of 1 to 3 people; each person can be in only one team.

Project ideas

Repair Safety Drift with SafeTune

Use SafeTune and AuditKIT to audit and repair safety drift in a small Cohere Labs model, reporting before-and-after safety metrics.

Architecture: CuratorKIT builds safety and red-team training data → AlignTune fine-tunes a small Cohere Labs model → SafeTune audits and repairs safety drift → AuditKIT compares before-and-after safety metrics

Stack: SafeTune, AlignTune, AuditKIT, Hugging Face

Improve Tool Calls with AgentTune

Post-train a small Cohere Labs model with AgentTune for more accurate, efficient, reliable tool calls, then measure gains with AuditKIT.

Architecture: Collect tool-call traces and success labels → AgentTune post-trains for accurate tool calls → AuditKIT benchmarks before-and-after reliability → Publish adapter, logs, and datasets on Hugging Face

Stack: AgentTune, AuditKIT, Hugging Face, Cohere Labs open-weights model

Circuit Interventions for Safety

Use CircuitKIT to find, evaluate, and intervene on safety-relevant circuits in a Cohere Labs model, reporting before-and-after metrics.

Architecture: Pick a safety behavior and collect evaluation prompts → CircuitKIT discovers and evaluates candidate circuits → Intervene on selected circuits and rerun AuditKIT → Publish circuit findings and metrics on Hugging Face

Stack: CircuitKIT, AuditKIT, Hugging Face, Cohere Labs open-weights model

FAQ

Who is eligible to enter?
The event is fully online and no blocked countries are listed, but eligibility is not confirmed on the event page. Read the official rules to confirm who can enter.

Sources

Open the original event page →

More open AI hackathons