knowledge base
Hackathon knowledge base
Working notes for the Lexsi Labs AI Safety & Alignment Hackathon, with Cohere and Hugging Face, November 2026. Each page builds one method from its smallest parts, with fixed definitions and small worked numbers.
What this is for: reading machine-learning papers about large language models (LLMs) is easier when every term has one fixed meaning. Every underlined term links to the glossary. Abbreviations are spelled out at first use on each page.
Pages
The steering code we are using (SafeTune)
SafeTune is Lexsi Labs' library for auditing and repairing safety behaviour in fine-tuned models. Its steering methods live in one folder. The one we use is Conditional Activation Steering (CAST), which adds a behaviour vector only when a condition vector fires. Its behaviour vector comes from Contrastive Activation Addition (CAA). Explanation: steering §10.
- CASTModel documentation pagesafetune.lexsi.ai · parameters, defaults and a full example
- src/safetune/interventions/steer/cast.pyThe implementation. Holds
fit_cast_condition(fits the condition vector and grid-searches layer and threshold) andCASTModel(applies the gated steering). - Exact lines at commit
3901a9f: fit_cast_condition (line 195) · CASTModel (line 411) - src/safetune/interventions/steer/caa.py · extract_caa_vectors (line 62) · CAAModel docsBuilds the behaviour vector that CAST adds.
- CASTTrainer in src/safetune/runner/steer/_activation.py (line 139)The higher-level wrapper:
steer.CASTTrainer(model, tokenizer).calibrate(...). - All steering methods (folder) · Steer overview in the docsAlso includes
refusal_direction.py(Arditi et al.) andprobe_guard.py(a linear probe gate).
Note: SafeTune's code and docs credit CAST to "Wu et al." The paper on arXiv (2409.05907) is by Bruce W. Lee, Inkit Padhi and colleagues at IBM Research. IBM's original library: IBM/activation-steering. Line numbers are for commit 3901a9f (3 November 2026); the main links follow later changes.
Lexsi Labs tools
- SafeTune · docsAuditing and repairing safety drift in fine-tuned models: steering, unlearning, hardening.
- AlignTune · docsModular toolkit for post-training alignment, such as preference training.
- AuditKIT · docsOne library for benchmark, judge, red-team, security and performance evaluation.
- CircuitKIT · docsDiscover, evaluate and intervene on circuits in transformer models.
- CuratorKIT · docsData curation and synthetic data generation for post-training.
- AgentTune · docsReinforcement-learning trainer for tool-calling agents, plus a build, collect, evaluate, train pipeline.
- DLBacktrace · docsAn explainability technique for deep-learning models of any type.
- All Lexsi Labs repositories
Hackathon and models
- hackathon.lexsi.aiThe event site: rules, dates, tracks.
- Aya Expanse 8B (
CohereLabs/aya-expanse-8b)Cohere Labs' multilingual chat model, 8 billion parameters. Access is gated: accept the terms on Hugging Face first. - Tiny Aya collection · tiny-aya-globalSmaller multilingual models, about 3.3 billion parameters. The collection also has base, regional and "thinker" variants. Gated.
Private project notes
Project notes (password protected): the journal, night reports, literature reviews, results and the definitions of every assistant behaviour.