Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add skills/zjunlp/mechanist/crosscodernpx skills add zjunlp/Mechanist --skill crosscodergit clone --depth 1 https://github.com/zjunlp/MechanistWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00070 | $0.05486 |
| Opus 5 | $0.00035 | $0.02743 |
| Sonnet 5 | $0.00014 | $0.01097 |
| Haiku 4.5 | $0.00007 | $0.00549 |
Grade A, and why
crosscoder-learning scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 642 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Dictionary Learning & Crosscoders Skill
When to Use
Activate this skill when the user needs to:
- Train Sparse Autoencoders (SAEs) on neural network activations (MLP outputs, attention outputs, residual streams)
- Train CrossCoders to compare activations across two related models (e.g., a base model and its fine-tuned variant)
- Load and use pretrained dictionaries (AutoEncoders, CrossCoders, JumpReLU SAEs) from local disk or the Hugging Face Hub
- Cache activations from a language model for later offline training
- Evaluate dictionaries using MSE loss, L1/L0 sparsity, CE diff, and variance-explained metrics
- Push or pull dictionary weights to/from the Hugging Face Hub
- Perform model diffing to identify which features change between a base model and a fine-tuned model
- Use
BatchTopKCrossCoderfor batch-level top-k sparsity across model pairs - Work with the
nnsightlibrary to hook into language model internals
Keywords: sparse autoencoder, SAE, crosscoder, dictionary learning, activation buffer, neural network features, mechanistic interpretability, model diffing, JumpReLU, GatedSAE, TopK SAE, nnsight, Pythia, Gemma, residual stream, MLP output, attention output
Quick Reference
| Resource | URL |
|---|---|
| Repository (fork) | https://github.com/jkminder/dictionary_learning |
| Original Repository | https://github.com/saprmarks/dictionary_learning |
| Anthropic CrossCoder Paper | https://transformer-circuits.pub/drafts/crosscoders/index.html#model-diffing |
| BatchTopKCrossCoder Paper | https://arxiv.org/pdf/2504.02922 |
| nnsight Documentation | https://nnsight.net/ |
| nnsight Walkthrough | https://nnsight.net/notebooks/tutorials/walkthrough/ |
| Pretrained Dictionaries (Pythia) | https://baulab.us/u/smarks/autoencoders/ |
| Pretrained CrossCoder (HF Hub) | Butanium/gemma-2-2b-crosscoder-l13-mu4.1e-02-lr1e-04 |
| Pretrained SAE Downloader Script | ./pretrained_dictionary_downloader.sh |
Installation / Setup
What ships with it
3 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 2d ago First seen · 642 lines · 70 tokens per session scan A e983ce79e9f0
crosscoder-learning is a skill published in the GitHub repository zjunlp/Mechanist (51 stars, last pushed 6d ago), licensed MIT. It adds 70 tokens to every session and 5,486 once invoked, about $0.0003 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
nanoresearch-writing
Draft a LaTeX research paper from all previous stage outputs.
nanoresearch-experiment
Generate a Python code skeleton from an experiment blueprint.
nanoresearch-ideation
Search academic literature and generate research hypotheses.
scienceworld-ambiguous-action-resolution
Use when the ScienceWorld environment returns an "Ambiguous request" prompt with a numbered list of identical action options. This skill resolves the disambiguation by selecting the lowest available number (typically 0) to proceed, ensuring task progression when multiple identical object instances exist and the system…
scienceworld-animal-identifier
Use when the agent needs to locate, identify, and focus on a specific animal or biological entity in the ScienceWorld environment. This skill handles tasks involving animal comparison, examination, or interaction (such as determining lifespan extremes) by navigating to the correct location with "teleport to"…
brainstorming-research-ideas
Guides researchers through structured ideation frameworks to discover high-impact research directions. Use when exploring new problem spaces, pivoting between projects, or seeking novel angles on existing work.