Skip to content
Back to search

autolaya

SkillWorks with Claude CodeGitHub

"Deploy AutoLaya and guide an end-to-end domain decision-model build: define Schema, prepare seeds/data, generate and review samples, freeze datasets, train Laya or GLiNER, evaluate, test in Playground, and deploy a local API. Use when operating, configuring, troubleshooting, or extending AutoLaya for a user."

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/dreamblooms--jev-autolaya/badge.svg)](/listings/dreamblooms--jev-autolaya)

HTML

<a href="/listings/dreamblooms--jev-autolaya"><img src="/api/listings/dreamblooms--jev-autolaya/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

Requirements:

- Python 3.12+ - Node.js 20.19+ (22+ recommended) - 16 GB RAM; 32 GB+ recommended for training - Optional GPU: NVIDIA CUDA or AMD ROCm; CPU works for Fit and inference

Clone the repository:

git clone https://github.com/DreamBlooms/Jev-AutoLaya.git
cd Jev-AutoLaya

Install and start:

./deploy.sh setup
./deploy.sh setup-laya
./deploy.sh setup-gliner   # only when using GLiNER
./deploy.sh start

Hand It to Your Agent

You can also give the repository to an agent. Point it to [SKILL.md](SKILL.md) and use a prompt like:

> Read `SKILL.md` first (https://raw.githubusercontent.com/DreamBlooms/Jev-AutoLaya/refs/heads/main/SKILL.md), then help me deploy AutoLaya and build a domain decision model from real examples through Schema, generation, review, dataset creation, training, evaluation, Playground testing, and deployment. Stop and ask before destructive changes or when you need more examples.

The skill contains operating rules, safety checks, the docs map, and the end-to-end workflow.

Open <http://localhost:8000>. API docs are at <http://localhost:8000/docs>.

The first visit opens a guided walkthrough using mock data. It never modifies real projects and can be replayed from Settings → Product guide.

Configure Keys

`deploy.sh` creates `backend/.env` automatically. Add your LLM and Jev keys there:

LLM_BASE_URL=https://api.deepseek.com
LLM_MODEL=deepseek-flash
LLM_API_KEY=your_key

# Jev through OpenRouter
JEV_PROVIDER=openrouter
OPENROUTER_API_KEY=your_key
OPENROUTER_JEV_MODEL=~typesafe/jev-latest

You can use any OpenAI-compatible Chat Completions service. Without Jev, choose direct LLM labels or no teacher.

Restart after editing:

./deploy.sh restart

Download Models

Laya and GLiNER base models:

MODEL_SOURCE=hf MODEL_SET=base ./deploy.sh pull-models
MODEL_SOURCE=hf-mirror MODEL_SET=base ./deploy.sh pull-models
MODEL_SOURCE=ms MODEL_SET=base ./deploy.sh pull-models

Laya Fit models:

FIT_SOURCE=hf ./deploy.sh pull-fit
FIT_SOURCE=hf-mirror ./deploy.sh pull-fit
FIT_SOURCE=hf FIT_BACKBONE_SOURCE=ms ./deploy.sh pull-fit

GPU

NVIDIA CUDA:

CUDA_TORCH_INDEX=https://download.pytorch.org/whl/cu128 ./deploy.sh setup-laya
CUDA_TORCH_INDEX=https://download.pytorch.org/whl/cu128 ./deploy.sh setup-gliner

For AMD ROCm or CPU, use the matching PyTorch wheel index:

LAYA_TORCH_INDEX=https://download.pytorch.org/whl/rocm6.4 ./deploy.sh setup-laya
LAYA_TORCH_INDEX=https://download.pytorch.org/whl/cpu ./deploy.sh setup-laya

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

SKILL.md · Manifest · 276 lines
  1. ---
  2. name: autolaya
  3. description: "Deploy AutoLaya and guide an end-to-end domain decision-model build: define Schema, prepare seeds/data, generate and review samples, freeze datasets, train Laya or GLiNER, evaluate, test in Playground, and deploy a local API. Use when operating, configuring, troubleshooting, or extending AutoLaya for a user."
  4. ---
  5. # AutoLaya Operator Skill
  6. Use this skill to operate AutoLaya safely and to guide a user from a domain description to a deployed decision model.
  7. ## Core Rule
  8. **Read the existing docs before explaining details or inventing commands.** Do not duplicate long UI instructions, API payloads, or troubleshooting tables in chat. Point to the relevant document and summarize only the decision the user needs.
  9. Docs map:
  10. | Need | Read |
  11. |---|---|
  12. | First install, keys, model download, first project | [docs/guide/getting-started.md](docs/guide/getting-started.md) |
  13. | Schema, seeds/import, generation, review, snapshots | [docs/guide/project-workflow.md](docs/guide/project-workflow.md) |
  14. | Laya/GLiNER training, evaluation, artifacts, Playground | [docs/guide/training-evaluation.md](docs/guide/training-evaluation.md) |
  15. | Serving, `/v1/systemone`, auth, backups | [docs/guide/deployment-api.md](docs/guide/deployment-api.md) |
  16. | User-facing failures and recovery | [docs/guide/troubleshooting.md](docs/guide/troubleshooting.md) |
  17. | Exact REST payloads and endpoints | [docs/API.md](docs/API.md) |
  18. | Current capabilities and limits | [docs/STATUS.md](docs/STATUS.md) |
  19. | Internals only when changing/debugging code | [docs/DESIGN.md](docs/DESIGN.md) |
  20. Chinese docs mirror these paths under `docs/guide/` and `docs/API_ZH.md`.
  21. ## Hand This Repository to Your Agent
  22. This repository is designed to be handed to an agent together with this `SKILL.md` file.
  23. Give the agent the repository and a prompt such as:
  24. > Read `SKILL.md` first (https://raw.githubusercontent.com/DreamBlooms/Jev-AutoLaya/refs/heads/main/SKILL.md), then help me deploy AutoLaya and build a domain decision model from real examples through Schema, generation, review, dataset creation, training, evaluation, Playground testing, and deployment. Stop and ask before destructive changes or when you need more examples.
  25. The agent should use this skill for operating rules, the docs map for details, and the user's real use case and samples for the actual project.
  26. ## Understand the Real Use Case First
  27. Before creating a project, Schema, or generation job, make sure you understand the user's actual use case:
  28. - What decision is being made, and what action follows from it?
  29. - Who consumes the answer, and what is the cost of a wrong answer?
  30. - What are the labels, score levels, and genuinely ambiguous cases?
  31. - What inputs occur in production: language, length, structure, noise, and frequency?
  32. - What latency, cost, and review capacity are acceptable?
  33. - Which real examples, existing labels, and held-out data already exist?
  34. If the user gives only a vague domain, ask for **5–20 real, anonymized examples** and a clear statement of how each should be judged. Prefer real examples over invented ones.
  35. **Data quality determines final model quality.** Synthetic data and teacher labels are starting points, not ground truth. Use human-confirmed labels, check label consistency and boundary cases, cover hard negatives, and keep real held-out data for final evaluation. Do not optimize for a fast training run at the cost of a dataset that does not represent the real use case.
  36. ## Safety Rules
  37. 1. Never print, copy, commit, or expose API keys, SSH passwords, `.env` contents, or user data.
  38. 2. Ask before destructive actions: deleting a project, snapshot, artifact, or deployment; overwriting `backend/.env`; resetting `data/`.
  39. 3. Before restarting AutoLaya, check for active jobs. Queued jobs survive a restart; running jobs may fail and need to be restarted.
  40. 4. Do not expose the management UI or API publicly. There is no built-in auth.
  41. 5. Do not claim a model, GPU path, or integration works unless it was verified in this environment.
  42. 6. Do not commit changes unless the user explicitly asks.
  43. ## First Contact
  44. Determine which mode the user needs:
  45. - **Existing deployment:** check health first; do not reinstall.
  46. - **Fresh local install:** follow `deploy.sh` setup.
  47. - **Remote deployment:** use the host, path, and credentials supplied by the user; preserve `.env`, `data/`, `models/`, and virtual environments.
  48. - **Build a model:** ask for the real decision, who uses it, 5–20 real examples if available, existing labels/held-out data, target model, and whether data already exists.
  49. ### Baseline Checks
  50. ```bash
  51. pwd
  52. ./deploy.sh status
  53. curl -s http://localhost:8000/api/health
  54. curl -s http://localhost:8000/api/v1/config
  55. ```
  56. If the host or port differs, use the user's actual URL.
  57. For active work:
  58. ```bash
  59. curl -s 'http://localhost:8000/api/v1/generation-jobs'
  60. curl -s 'http://localhost:8000/api/v1/training-jobs'
  61. curl -s 'http://localhost:8000/api/v1/evaluations'
  62. ```
  63. ## Deployment Playbook
  64. Only deploy what is needed. Do not reinstall runtimes when they are already ready.
  65. ```bash
  66. ./deploy.sh setup
  67. ./deploy.sh setup-laya
  68. ./deploy.sh setup-gliner # only when GLiNER is required
  69. ./deploy.sh start
  70. ```
  71. Model downloads:
  72. ```bash
  73. MODEL_SOURCE=hf MODEL_SET=base ./deploy.sh pull-models
  74. FIT_SOURCE=hf ./deploy.sh pull-fit
  75. ```
  76. GPU install uses the matching PyTorch wheel index. See the quick-start guide rather than guessing a wheel URL.
  77. After deployment:
  78. ```bash
  79. ./deploy.sh status
  80. curl -s http://localhost:8000/api/health
  81. curl -s http://localhost:8000/api/v1/config
  82. ```
  83. Confirm `llmConfigured`, `jevConfigured`, `availableDevices`, `layaRuntimeReady`, and `glinerRuntimeReady` before starting a model workflow.
  84. ### Remote Sync
  85. - Preserve `backend/.env`, `data/`, `models/`, `backend/.venv`, `backend/laya-venv`, and `backend/gliner-venv`.
  86. - Sync application code and built assets, restart, then repeat health/config checks.
  87. - Do not copy local secrets to a remote host unless the user asks.
  88. - Do not run a remote restart while jobs are active without confirmation.
  89. ## Full Build Checklist
  90. Use the UI or API. Prefer the API when the user wants repeatable automation.
  91. ### 1. Project and Schema
  92. - Create one project for one domain.
  93. - Add at least one `Choice`, `Noul`, or `Score` question.
  94. - Validate keys are unique and options are meaningful.
  95. - Detail: [project workflow](docs/guide/project-workflow.md).
  96. ### 2. Seeds and Samples
  97. - Add 5–20 real seeds when possible.
  98. - Use JSONL/CSV templates for bulk input.
  99. - Imported rows remain `pending`.
  100. - Real held-out data must be marked as held out.
  101. - Detail: [seeds and import](docs/guide/project-workflow.md).
  102. ### 3. Generate Data
  103. - Choose Jev Teacher, direct LLM labels, or no teacher.
  104. - Set count, languages, difficulty mix, negative ratio, and label balancing.
  105. - Advanced options include custom prompt, Think budget, Think level, and temperature.
  106. - Preview the LLM request before launching a large job.
  107. - Detail: [generation](docs/guide/project-workflow.md).
  108. ### 4. Human Review
  109. - Every question needs an answer before acceptance.
  110. - Use batch accept only after spot-checking quality.
  111. - Reject bad samples with a reason.
  112. - Do not skip review for production data.
  113. - Detail: [review](docs/guide/project-workflow.md).
  114. ### 5. Freeze Dataset
  115. - Confirm accepted samples are complete.
  116. - Freeze a snapshot and verify split counts and label distribution.
  117. - Require real held-out data for formal evaluation.
  118. - Detail: [freeze a dataset](docs/guide/project-workflow.md).
  119. ### 6. Train
  120. - Start with Laya Fit for a baseline.
  121. - Fine-tune Laya when data and hardware justify it.
  122. - GLiNER is fine-tune only.
  123. - Confirm the runtime and device before launching.
  124. - Detail: [training](docs/guide/training-evaluation.md).
  125. ### 7. Evaluate
  126. - Evaluate on validation/test rows, not training rows.
  127. - Compare Base, Fit, fine-tuned, and optionally Jev.
  128. - Use “More” to select historical trained versions.
  129. - Prefer held-out results over training loss.
  130. - Detail: [evaluation](docs/guide/training-evaluation.md).
  131. ### 8. Test in Playground
  132. - Use Playground for single-input checks without creating datasets or jobs.
  133. - Use project Schema, presets, or a custom editable Schema.
  134. - State may be a string, object, or array.
  135. - Export a SystemOne cURL when the user needs an integration example.
  136. - Detail: [Playground](docs/guide/training-evaluation.md).
  137. ### 9. Deploy and Verify
  138. - Deploy the selected artifact.
  139. - Confirm `/health`.
  140. - Call `/v1/systemone` with the same questions used during training.
  141. - For GLiNER, call `/v1/classify` or the compatible mapping.
  142. - Detail: [deployment](docs/guide/deployment-api.md).
  143. ## API Quick Reference
  144. Base URL:
  145. ```text
  146. http://localhost:8000/api/v1
  147. ```
  148. Use [docs/API.md](docs/API.md) for request bodies and response fields. The most common workflow endpoints are:
  149. ```text
  150. POST /projects
  151. POST /projects/{id}/questions
  152. POST /projects/{id}/seeds
  153. POST /projects/{id}/samples/import
  154. POST /projects/{id}/generate
  155. GET /generation-jobs/{id}/events
  156. POST /samples/{id}/accept
  157. POST /projects/{id}/datasets
  158. POST /training-jobs
  159. GET /training-jobs/{id}/events
  160. POST /evaluations
  161. POST /playground/run
  162. POST /deployments
  163. ```
  164. Rules:
  165. - Use the API for repeatable automation.
  166. - Do not poll aggressively; use SSE event endpoints for long jobs.
  167. - Playground and base-model loading can take 10–60 seconds.
  168. - Treat model execution as serialized; do not launch overlapping GPU-heavy runs.
  169. ## Decision Guide
  170. ### “Deploy this for me”
  171. -> Follow the deployment playbook, verify health/config, stop before creating data unless asked.
  172. ### “Build a model from this domain”
  173. -> Get the domain and schema first. Then project → Schema → seeds → generation → review → snapshot → Fit → evaluation → deploy.
  174. ### “Test this text/model”
  175. -> Use Playground. Do not create training data just to test one input.
  176. ### “Why is training/evaluation bad?”
  177. -> Check data size, label quality, held-out leakage, selected base, and whether the correct snapshot/artifact was used. Read evaluation output before changing parameters.
  178. ### “Add/fix a feature”
  179. -> Read [DESIGN.md](docs/DESIGN.md) only for internals. Inspect current routes and tests. Run:
  180. ```bash
  181. cd backend && .venv/bin/ruff check . && .venv/bin/pytest
  182. cd frontend && npm run typecheck && npm run build
  183. ```
  184. ## Completion Criteria
  185. Before telling the user the workflow is complete, verify:
  186. - the real use case and decision boundary are clear;
  187. - project has a valid Schema;
  188. - real examples and held-out data were used where possible;
  189. - generated samples and labels were reviewed and confirmed;
  190. - snapshot exists and contains held-out data;
  191. - training job completed and an artifact exists;
  192. - evaluation used validation/test rows;
  193. - deployment health check passes;
  194. - a real inference request returns answers;
  195. - no secrets or user data were exposed.
README.md · README · 181 lines
  1. # AutoLaya
  2. [简体中文](README_ZH.md) | English
  3. **AutoLaya** is a local-first tool for building domain decision models. It turns a domain description and a few real examples into a Laya or GLiNER model that you can train, evaluate, and deploy yourself.
  4. <img width="1920" height="980" alt="img-welcome" src="https://github.com/user-attachments/assets/f63323a5-e757-4aee-a820-2f9b48d2a4f2" />
  5. ```text
  6. domain description / seed examples
  7. → LLM-generated domain text
  8. → Jev / LLM / human labels
  9. → human review
  10. → frozen dataset
  11. → train Laya / GLiNER
  12. → evaluate on held-out rows
  13. → deploy a local decision API
  14. ```
  15. <img width="1920" height="980" alt="img-s1" src="https://github.com/user-attachments/assets/5432c848-f52b-4a15-aaf4-a36709e6e70f" />
  16. Everything happens in one web workspace. You do not need to write training scripts or send business data to a managed training platform.
  17. ## What You Can Do
  18. - **Define the domain and Schema:** use Choice, Noul, and Score questions.
  19. - **Prepare data:** add examples manually, import JSONL/CSV, generate with an LLM, or import real held-out data.
  20. - **Label automatically:** use Jev Teacher, direct LLM weak labels, or fully manual labels.
  21. - **Review:** edit, accept, reject, and batch-process samples.
  22. - **Manage datasets:** freeze immutable snapshots and download Canonical, Laya, or GLiNER JSONL.
  23. - **Train models:** Laya Fit + calibration, Laya fine-tuning, and GLiNER2.5 fine-tuning.
  24. - **Evaluate models:** compare Base, Fit, and fine-tuned models, including historical training versions.
  25. - **Test models:** edit Schema and State freely in the Playground and compare models.
  26. - **Deploy models:** start local services and call the Jev-compatible `/v1/systemone`.
  27. - **Export requests:** generate a complete SystemOne cURL request from the Playground.
  28. ## Quick Start
  29. Requirements:
  30. - Python 3.12+
  31. - Node.js 20.19+ (22+ recommended)
  32. - 16 GB RAM; 32 GB+ recommended for training
  33. - Optional GPU: NVIDIA CUDA or AMD ROCm; CPU works for Fit and inference
  34. Clone the repository:
  35. ```bash
  36. git clone https://github.com/DreamBlooms/Jev-AutoLaya.git
  37. cd Jev-AutoLaya
  38. ```
  39. Install and start:
  40. ```bash
  41. ./deploy.sh setup
  42. ./deploy.sh setup-laya
  43. ./deploy.sh setup-gliner # only when using GLiNER
  44. ./deploy.sh start
  45. ```
  46. ### Hand It to Your Agent
  47. You can also give the repository to an agent. Point it to [SKILL.md](SKILL.md) and use a prompt like:
  48. > Read `SKILL.md` first (https://raw.githubusercontent.com/DreamBlooms/Jev-AutoLaya/refs/heads/main/SKILL.md), then help me deploy AutoLaya and build a domain decision model from real examples through Schema, generation, review, dataset creation, training, evaluation, Playground testing, and deployment. Stop and ask before destructive changes or when you need more examples.
  49. The skill contains operating rules, safety checks, the docs map, and the end-to-end workflow.
  50. Open <http://localhost:8000>. API docs are at <http://localhost:8000/docs>.
  51. The first visit opens a guided walkthrough using mock data. It never modifies real projects and can be replayed from Settings → Product guide.
  52. ### Configure Keys
  53. `deploy.sh` creates `backend/.env` automatically. Add your LLM and Jev keys there:
  54. ```env
  55. LLM_BASE_URL=https://api.deepseek.com
  56. LLM_MODEL=deepseek-flash
  57. LLM_API_KEY=your_key
  58. # Jev through OpenRouter
  59. JEV_PROVIDER=openrouter
  60. OPENROUTER_API_KEY=your_key
  61. OPENROUTER_JEV_MODEL=~typesafe/jev-latest
  62. ```
  63. You can use any OpenAI-compatible Chat Completions service. Without Jev, choose direct LLM labels or no teacher.
  64. Restart after editing:
  65. ```bash
  66. ./deploy.sh restart
  67. ```
  68. ### Download Models
  69. Laya and GLiNER base models:
  70. ```bash
  71. MODEL_SOURCE=hf MODEL_SET=base ./deploy.sh pull-models
  72. MODEL_SOURCE=hf-mirror MODEL_SET=base ./deploy.sh pull-models
  73. MODEL_SOURCE=ms MODEL_SET=base ./deploy.sh pull-models
  74. ```
  75. Laya Fit models:
  76. ```bash
  77. FIT_SOURCE=hf ./deploy.sh pull-fit
  78. FIT_SOURCE=hf-mirror ./deploy.sh pull-fit
  79. FIT_SOURCE=hf FIT_BACKBONE_SOURCE=ms ./deploy.sh pull-fit
  80. ```
  81. ### GPU
  82. NVIDIA CUDA:
  83. ```bash
  84. CUDA_TORCH_INDEX=https://download.pytorch.org/whl/cu128 ./deploy.sh setup-laya
  85. CUDA_TORCH_INDEX=https://download.pytorch.org/whl/cu128 ./deploy.sh setup-gliner
  86. ```
  87. For AMD ROCm or CPU, use the matching PyTorch wheel index:
  88. ```bash
  89. LAYA_TORCH_INDEX=https://download.pytorch.org/whl/rocm6.4 ./deploy.sh setup-laya
  90. LAYA_TORCH_INDEX=https://download.pytorch.org/whl/cpu ./deploy.sh setup-laya
  91. ```
  92. ## First Use
  93. 1. Open AutoLaya and finish the first-visit guide.
  94. 2. Create a project and choose its languages.
  95. 3. Add Choice, Noul, or Score questions in Schema.
  96. 4. Add seeds or import JSONL/CSV from a template.
  97. 5. Generate data and review the results.
  98. 6. Freeze a dataset.
  99. 7. Train Laya or GLiNER.
  100. 8. Compare models in Evaluation and test inputs in the Playground.
  101. 9. Deploy from Models to get a local `/v1/systemone` URL.
  102. See [Getting Started](docs/guide/en/getting-started.md) for the full walkthrough.
  103. ## Documentation
  104. | Document | Contents |
  105. |---|---|
  106. | [User Guide](docs/guide/en/README.md) | Complete usage guide |
  107. | [Getting Started](docs/guide/en/getting-started.md) | Install, configure, start, and build the first project |
  108. | [Project & Data Workflow](docs/guide/en/project-workflow.md) | Schema, seeds, generation, review, datasets |
  109. | [Training & Evaluation](docs/guide/en/training-evaluation.md) | Laya, GLiNER, evaluation, artifacts, Playground |
  110. | [Deployment & API](docs/guide/en/deployment-api.md) | Deploy, call endpoints, backup, security |
  111. | [Troubleshooting](docs/guide/en/troubleshooting.md) | Common problems and fixes |
  112. | [API Reference](docs/API.md) | Complete interface reference for integrations |
  113. | [Design](docs/DESIGN.md) | Architecture and design record |
  114. ## Operations
  115. ```bash
  116. ./deploy.sh start
  117. ./deploy.sh status
  118. ./deploy.sh logs
  119. ./deploy.sh restart
  120. ./deploy.sh stop
  121. ```
  122. Back up at least:
  123. - `backend/.env`
  124. - `data/autolaya.db`
  125. - `data/datasets/`
  126. - `data/artifacts/`
  127. - `models/` if you do not want to download them again
  128. ## Known Limitations
  129. - The management API has no built-in login or authentication; do not expose it publicly.
  130. - Multi-label decisions are not supported; use multiple Noul questions instead.
  131. - GLiNER provides labels and confidence, not full calibrated probabilities.

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.