Jeden Skill in Manus ausführen
mit einem Klick

Jeden Skill in Manus mit einem Klick ausführen

model-incorporation-metadata

Sterne2

Forks0

Aktualisiert1. Juli 2026 um 09:08

Fills in the required metadata fields for an Ersilia Model Hub model, given the metadata.yml file and the original publication PDF. Reads the source code URL directly from the metadata file — no need to provide it separately. Use this skill whenever a user wants to populate, complete, or update a metadata.yml file for an Ersilia model, mentions an Ersilia model contribution, or is working on model metadata for the Ersilia Model Hub. Trigger even if the user just says "fill in the metadata" or "help me with the metadata.yml" in any Ersilia context.

Installation

Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.

In Manus ausführen

Quelle

ersilia-os

ersilia-os/ersilia-skills

GitHub-Repository öffnen Creator-Repositorys ansehen

Download

In Manus ausführen

Verwandte BerufeSOC

Basierend auf der SOC-Berufsklassifikation

SoftwareentwicklerInformatik- und Mathematikberufe·SOC 15-1252

Datei-Explorer

2 Dateien

SKILL.md

readonly

Mehr aus diesem Repository

gleiches Repository

model-incorporation-code

ersilia-os/ersilia-skills

Integrate the code of a new ML model into an Ersilia model template repository. Use this skill whenever the user has: (1) a source model repository with the original ML code, (2) an ersilia-model-template repository (already forked/cloned), and (3) optionally a PDF of the scientific article — and needs to wire up the actual model code. The skill handles all coding steps: copying checkpoints with git-lfs tracking, adapting main.py to replace the molecular-weight placeholder with real inference, creating run_columns.csv, updating install.yml with pinned versions, and producing run_input.csv and run_output.csv by actually running the model. Trigger on phrases like "incorporate model code", "fill in the template", "adapt main.py", "add checkpoints", "create run_columns", "create install.yml for ersilia", "generate example files for ersilia model", or any request to wire a source model into the eos-template format.

2026-07-012

ersilia-model-test

ersilia-os/ersilia-skills

Tests an Ersilia Model Hub model before hub incorporation. Runs `ersilia test --shallow` on a locally cloned model repository, reads the results, identifies failing checks, and proposes a debugging strategy to the user — but does not apply fixes automatically. Use this skill whenever a user wants to test, validate, or debug an Ersilia model before submitting it to the hub, mentions test failures, says "the test is failing", or is preparing a model for hub incorporation. Trigger even if the user just says "test the model" or "run ersilia test" in any Ersilia context.

2026-07-012

model-incorporation-request

ersilia-os/ersilia-skills

Open a model request issue on ersilia-os/ersilia for a new model to be incorporated into the Ersilia Model Hub. Use this skill whenever a user wants to request a new model, open a model request, or start the model incorporation pipeline. The skill reads the publication PDF and source code repository to propose all required fields, asks the user to confirm, then opens the GitHub issue mimicking the official ersilia-os/ersilia model request form. Trigger on phrases like "request a model", "open a model request", "add a new model to the hub", "start model incorporation", or "open an issue for a new model".

2026-07-012

model-incorporation-reproduce

ersilia-os/ersilia-skills

Reproduce the key performance metrics reported in a paper for an Ersilia Model Hub model. Reads the publication PDF, extracts reported results (AUC, RMSE, accuracy, R², etc.), assesses whether each benchmark dataset is publicly available, acquires the datasets, runs the incorporated model, and compares reproduced vs reported metrics with tolerance-aware verdicts. Use this skill whenever a user wants to verify that an incorporated model performs as described in its paper, after running model-incorporation-code. Trigger on phrases like "reproduce the results", "check model performance", "validate the model against the paper", "does the model match what the paper reports", "benchmark the model", or any request to verify an incorporated Ersilia model against its publication.

2026-06-302

literature-digest

ersilia-os/ersilia-skills

Produce the weekly Ersilia literature digest. Primary mission: surface every model and dataset that could plausibly join the Ersilia Model Hub — across all six Hub task families (Property prediction, Activity prediction, Featurization, Projection, Similarity search, Generation), with antimicrobial / antipathogen drug discovery as the most-prioritised disease lens. Secondary mission: keep the team current on agentic AI for science, automation in chemistry, open-source drug discovery, and global health. Apply an explicit LMIC and decolonisation lens throughout. Use this skill whenever the user asks to prepare, run, or refresh the literature digest. Triggers include: "weekly literature digest", "literature digest for Ersilia", "/literature-digest", "lit digest this week", "what did we miss last week", "digest the literature". Always use this skill for digest requests even if the ask seems simple.

2026-06-302

literature-review

ersilia-os/ersilia-skills

Conduct deep, structured literature research for Ersilia drug discovery work. Use this skill whenever a user asks to find, review, or summarise scientific papers about a disease, biological target, compound class, or AI/ML method in the context of drug discovery — especially for neglected tropical diseases (NTDs), antimicrobial resistance (AMR), or global health. Triggers include target names (e.g. Mtb, PfDHFR, TryR, InhA), disease names (e.g. malaria, leishmaniasis, tuberculosis, Chagas, schistosomiasis), model or method types (e.g. GNN, QSAR, ChemBERTa), and requests like "find papers on", "what does the literature say about", "summarise research on", "literature review for", or "what's been published about". Always use this skill for Ersilia-context literature work, even if the request is phrased casually.

2026-06-162

name

model-incorporation-metadata

description

Ersilia Model Metadata Filler

Your job is to fill in specific fields of an Ersilia model's metadata.yml file using information extracted from the original publication (PDF) and the original source code repository.

What you receive

The user will provide:

A path to the metadata.yml file (already partially filled from the model request)
A path or URL to the original publication PDF

The source code repository URL and publication URL are already in metadata.yml — read them from the file in Step 1 rather than asking the user. Only ask if those fields are empty or still contain placeholder text.

What you must fill in

Fill only these fields — leave everything else exactly as it is:

Field	Type	Accepted values
`Deployment`	list	see `deployment.txt`
`Source`	string	see `source.txt`
`Source Type`	string	see `source_type.txt`
`Task`	string	see `task.txt`
`Subtask`	string	see `subtask.txt`
`Output`	list	see `output.txt`
`Output Dimension`	integer	number of output values per input compound
`Output Consistency`	string	see `output_consistency.txt`
`Interpretation`	string	one sentence, max 20 words
`Tag`	list	see `tag.txt`
`Biomedical Area`	list	see `biomedical_area.txt`
`Target Organism`	list	see `target_organism.txt`
`Publication`	string	must be a DOI URL: `https://doi.org/...`
`Publication Type`	string	see `publication_type.txt`
`Publication Year`	integer	year

Never modify: Identifier, Slug, Status, Title, Description, Source Code, License, Contributor, and any auto-populated fields (Incorporation Date, S3, DockerHub, Model Size, etc.)

Step-by-step workflow

0. Load allowed values from GitHub

Before filling anything, fetch all eleven controlled-vocabulary files directly from GitHub:

https://raw.githubusercontent.com/ersilia-os/ersilia/master/ersilia/hub/content/metadata/tag.txt
https://raw.githubusercontent.com/ersilia-os/ersilia/master/ersilia/hub/content/metadata/biomedical_area.txt
https://raw.githubusercontent.com/ersilia-os/ersilia/master/ersilia/hub/content/metadata/target_organism.txt
https://raw.githubusercontent.com/ersilia-os/ersilia/master/ersilia/hub/content/metadata/task.txt
https://raw.githubusercontent.com/ersilia-os/ersilia/master/ersilia/hub/content/metadata/subtask.txt
https://raw.githubusercontent.com/ersilia-os/ersilia/master/ersilia/hub/content/metadata/output.txt
https://raw.githubusercontent.com/ersilia-os/ersilia/master/ersilia/hub/content/metadata/output_consistency.txt
https://raw.githubusercontent.com/ersilia-os/ersilia/master/ersilia/hub/content/metadata/source.txt
https://raw.githubusercontent.com/ersilia-os/ersilia/master/ersilia/hub/content/metadata/source_type.txt
https://raw.githubusercontent.com/ersilia-os/ersilia/master/ersilia/hub/content/metadata/deployment.txt
https://raw.githubusercontent.com/ersilia-os/ersilia/master/ersilia/hub/content/metadata/publication_type.txt

Each file is newline-separated. Fetch all eleven now and keep the lists in memory — you will validate every value you write against them.

1. Read the existing metadata.yml

Note the Source Code URL and Publication URL — you will need them. Confirm which fields still need filling (they may have template placeholder text like "Biomedical Area 1" or multiple values comma-separated).

2. Read the publication

Use the PDF reading tools to extract:

What the model predicts (property, activity, representation, etc.)
Number of endpoints/outputs (crucial for Output Dimension)
Training dataset and target organisms
Disease or application area
Whether outputs are probabilities, measured values, or generated structures
Year of publication and publication venue (journal vs preprint)

3. Fetch the source code repository

Use WebFetch on the remote GitHub repository (the Source Code URL from metadata.yml). Do not read any local template files — run_output.csv, run_columns.csv, and model/framework/ do not exist yet at this stage of the pipeline. Fetch only remote content:

The repository README
The repository file tree (to understand project structure)
Specific remote source files only if needed to clarify output dimension or deployment type

From this, determine:

Number of model endpoints / output columns (cross-check with paper)
Whether the model calls an external API or runs locally
Whether the license/code is from external authors, the Ersilia team, or re-trained

4. Fill in each field

Work through the fields systematically:

Deployment + Source Use only values from the deployment.txt and source.txt lists fetched in Step 0.

Deployment: [Local] and Source: Local for the vast majority of models — the model runs in Ersilia's infrastructure.
Use Online only if the model posts predictions to an external third-party server/API that Ersilia does not control.
Deployment is a list; Source is a single string.

Source Type Use only values from the source_type.txt list fetched in Step 0.

External: the model was developed by third-party authors (most models incorporated from published papers).
Internal: developed by the Ersilia team themselves.
Replicated: Ersilia re-trained the model following the original authors' methodology.

Task + Subtask Use only values from the task.txt and subtask.txt lists fetched in Step 0. Choose Task first, then the corresponding Subtask:

Annotation → model assigns a label or score to a molecule
- Activity prediction if the output is biological/pharmacological activity
- Property calculation or prediction if the output is a physicochemical or ADMET property
Representation → model encodes a molecule into a numerical vector or projection
- Featurization if it produces a fixed-length embedding/descriptor vector
- Projection if it projects molecules into 2D/3D space
Sampling → model generates or retrieves molecules
- Generation if it generates new molecules
- Similarity search if it retrieves similar molecules from a database

Output Use only values from the output.txt list fetched in Step 0. Can be a list if the model has mixed output types.

Score: a probability or likelihood (0–1 range, binary classification output)
Value: a numerical measurement (IC50, logP, pKa, molecular weight, descriptors, embeddings…)
Compound: a generated or retrieved molecule (SMILES or InChI)
Text: natural language output
Example: a model returning both a probability and an associated value would be [Score, Value].

Output Dimension The number of output values produced per input compound. Only count continuous numeric outputs (scores, values) — do not count binary class labels separately. So a model with 6 endpoints each returning one probability score has Output Dimension 6, not 12.

This is often explicit in the paper ("6 endpoints", "512-dimensional vector"). If not stated directly:

Count the number of prediction tasks/endpoints described
Check the repository: look at model output shapes, column headers in example outputs, or README tables
For embedding models, look for the vector size (e.g., 512, 1024, 2048)
If the vector size is configurable (e.g., n_components is a user parameter), ask the user — do not guess a default from examples in the docs

Output Consistency Use only values from the output_consistency.txt list fetched in Step 0.

Fixed: the model always returns the same output for the same input (most QSAR models, classifiers, regression models)
Variable: the model is stochastic and may return different outputs on repeated runs (generative models, models with dropout at inference, sampling-based methods)

Interpretation Write exactly one sentence. Hard limit: 20 words. Do not write two sentences. Do not start with "The model" — start with the output itself.

Never use a colon (:) in the interpretation text — colons break YAML parsing. Use a semicolon (;) or rephrase instead.

Good examples:

Higher score indicates greater predicted probability of anti-malarial activity.
100 features encoding molecular structure from a pretrained MACAW autoencoder.
Predicted probability of AMES mutagenicity; values closer to 1 indicate higher risk.
Binary indicators (1 = substructure present, 0 = absent) and total hit count.

Tag Choose one or more values from the list you read in Step 0 (tag.txt). Every value you write must appear verbatim in that file. If no tag fits well, pick the closest match — do not invent new tags.

Biomedical Area Choose one or more values from the list you read in Step 0 (biomedical_area.txt). Every value must appear verbatim in that file. Use Any only if the model is truly domain-agnostic (e.g., a general-purpose molecular featurizer with no disease focus).

Target Organism Choose one or more values from the list you read in Step 0 (target_organism.txt). Every value must appear verbatim in that file. Use Any if the model is not organism-specific.

Publication The Publication field must be a DOI URL in the format https://doi.org/.... This applies to both peer-reviewed articles and preprints (bioRxiv, ChemRxiv, and arXiv all issue DOIs). If the existing metadata contains a PubMed URL, abstract page, or journal landing page, locate and substitute the DOI. If genuinely no DOI exists, use the most stable URL available and note this to the user.

Publication Type Use only values from the publication_type.txt list fetched in Step 0.

Peer reviewed for articles published in journals (DOI typically starts with 10.1, 10.3, 10.7, etc.)
Preprint for bioRxiv (10.1101/), ChemRxiv (10.26434/), arXiv (10.48550/)
Other only in exceptional cases (thesis, technical report, no publication)

Publication Year Extract the year of publication from the paper or publication URL.

5. Propose all changes and ask for confirmation

Before writing anything, present every proposed value in a table:

Field	Current value	Proposed value	Source
Deployment	…	[Local]	default
Source	…	Local	default
Task	…	Annotation	paper §2
…	…	…	…

The "Source" column should briefly indicate where the value came from (e.g. "paper §2", "repo README", "default", "user input"). Flag any field where the evidence was ambiguous with a note in the Source column (e.g. "paper unclear — assumed 512").

Then ask:

"These are the values I propose to write to metadata.yml. Please confirm or correct any field before I update the file."

Only proceed to Step 6 once the user confirms. If the user corrects a value, update the table and re-show it before asking again.

6. Write the updated metadata.yml

Edit the file in place, replacing only the confirmed fields. Keep the YAML formatting consistent with the rest of the file (use list syntax for list fields, plain string for string fields, integer for integer fields).

YAML safety: any string value that contains a colon (:) must be wrapped in double quotes, e.g. Interpretation: "Predicted log solubility: higher values indicate better solubility." — otherwise the YAML is invalid. Before writing, check every string field you are changing for colons and quote it if needed.

7. Confirm and hand off

Print a one-line confirmation: "Done — N fields updated in <path>/metadata.yml."

Then list the remaining pipeline steps:

Remaining steps:

Run /model-incorporation-code to wire up the model code into the template repository

Run /ersilia-model-test to validate the model runs correctly

Run /model-incorporation-reproduce to verify the model performs as reported in the paper

Push and open a pull request from your fork to ersilia-os/eosXXXX

When you cannot determine a value

If you've read the paper and the repository and still cannot confidently determine a value (especially Output Dimension), state clearly what you found and what is unclear, and ask the user to clarify. Do not guess or leave placeholder text.