Blame for README.md

---
license: cc-by-4.0
language:
- en
pretty_name: "AI App Cloud Deploy Issues (GitHub)"
size_categories:
- 1K<n<10K
task_categories:
- text-classification
tags:
- deployment
- llm-apps
- devops
- github-issues
- ai-generated-code
configs:
- config_name: default
  data_files:
  - split: train
    path: data/threads.csv
---

# AI App Cloud Deploy Issues (GitHub)

**What this is:** a coded index of public GitHub issue threads where people try to get an **AI app** (a chatbot, a RAG app, an agent, a model server, or an app generated by an AI builder) running on a **remote host** (Vercel, Azure, AWS, Kubernetes, Hugging Face Spaces, Railway…).
Each row is one thread: a link, its metadata (dates, counts, state), and columns we coded from the text: what kind of thread it is (problem, feature request, docs request, question, CI), host, app type, framework, model provider, topic (what the thread is about: a failure, a feature or a question), error type, how it ended.

**Who made it:** Hal9. We build Cloudish, a hosting service for apps that AI agents deploy, so we have an interest in this topic. The dataset is meant to be useful on its own: it does not mention Cloudish, and Cloudish is not one of the hosts in it.

**No text, no authors.** The files contain **no issue text, no titles, no excerpts and no usernames** (not even hashed). The text belongs to its authors and stays on GitHub: follow the `url` to read a thread.

## Quick start

```python
from datasets import load_dataset

ds = load_dataset("hal9/ai-app-cloud-deploy-issues", split="train")  # repo id is a placeholder
df = ds.to_pandas()
print(df.groupby("target_host")["resolved"].apply(lambda s: (s == "yes").mean()).sort_values())
```

## Columns

See [`codebook.md`](codebook.md) for every column, its values and the rule behind it. In short:

- **metadata** (from the GitHub API): `url`, `repo`, `repo_group`, `created_at`, `closed_at`, `state`, `state_reason`, `n_comments`, `n_reactions`, `time_to_first_response_h`, `time_to_close_h`, `maintainer_replied`
- **coded**: `thread_type`, `target_host`, `app_type`, `framework`, `stack`, `model_provider`, `topic`, `error_type`, `resolution_type`, `resolved`, `ai_generated_app`, `ai_tool`
- **provenance**: `coding_method` (`keyword` or `keyword+agent_reviewed`), `coding_confidence` (`title`, `body`, `none`: where the error or topic keywords were found), `relevance_method` (`keyword` or `agent_reviewed`), `collected_on`

## How it was built

1. **Sources.** Public GitHub issues only, created since 2024-01-01, from 34 repositories chosen for precision: AI app frameworks and apps (Gradio, Chainlit, LangGraph, Vercel AI SDK, Flowise, Langflow, Open WebUI, Dify, TGI, vLLM, Ollama…), cloud vendors' AI sample apps (Azure, Google Cloud), and AI app builders / coding agents (bolt.diy, Dyad, Claude Code). The list is in `scripts/repos_full.json`.
   In a pilot, issue trackers of deploy CLIs (Vercel, Netlify, Fly, Railway) gave **0%** relevant threads about AI apps, and LangChain (Python) and LlamaIndex about 21%: they were dropped. SkyPilot (GPU jobs rather than app deploys) and replicate/cog (2 threads) were dropped after the full-build review.
2. **Search.** One GitHub search query per (repository, title term), for example `repo:Chainlit/chainlit is:issue "cloud run" in:title`. The terms are deploy words (`deploy`, `production`, `hosting`) and host names.
3. **Caps.** At most 2 threads per author across the whole dataset (the author is used for this cap during collection, then dropped), and at most 250 threads per repository.
4. **Relevance filter.** Nine written rules decide what "deploying or running an AI app on a host" means (see the codebook): a host is a remote machine (not a laptop, WSL or a local GPU); model serving on a host counts; an AI agent that deploys and hits its own limits counts; an app bug that merely shows up on a server does not; calling a cloud's AI API without deploying an app does not; in general frameworks (Streamlit, Gradio) the thread itself must show an AI app; CI of a cloud integration counts. Keyword rules apply them; every row the keywords could not settle (about 220) was read and judged by the AI assistant. Feature and docs requests are kept and labelled by `thread_type`.
5. **Coding.** Keyword rules (`scripts/code.py`), with no paid API. A sample of threads was then read and coded by the AI assistant (Claude) that built the dataset; its codes replace the keyword codes for those rows (`coding_method = keyword+agent_reviewed`). The relevance rules were set with a person from our team, who also checked a small sample of rows against the threads; the quality figures below come from the larger samples the AI reviewer read.

## Quality (measured, not assumed)

On samples read thread by thread by the reviewer (see step 5):

- **relevance precision** of rows decided by keywords alone (`relevance_method = keyword`): **70%** (42 of a fresh random sample of 60, drawn after the last rule change; 95% Wilson interval 58–80%). Counting the rows the reviewer read (`agent_reviewed`, about 28% of the file) as correct, the whole file is about **79%** relevant (about 70–86%). Most false positives are an app bug or an API call that mentions a deployment in passing, or a local install. These figures come from the AI reviewer, who also wrote the keyword rules
- **thread type** (`thread_type`) agrees with the reviewer on 49 of the same 60 rows (82%); the keyword `other` is the weak value
- **agreement of keyword codes with reviewed codes**: host 0.95–0.98, app type 0.95, framework 0.95–0.98, model provider 0.83–0.85, topic 0.48–0.70, error type 0.57–0.80, resolution 0.65–0.85 (two samples of 46 and 20 threads; the higher figure comes from a sample where the reviewer could see the keyword codes, so it is probably optimistic). The topic and error type figures were measured before three topics (`agent_behavior`, `observability`, `performance_scaling`) and four `agent_*` error types were added; those only relabel rows that were `other`, and their precision was checked separately: 28/28, 17/18 and 17/19 rows read.

`topic` and `error_type` are the weakest columns. Filter on `coding_confidence == "title"` when you need them to be reliable.

## Limits and biases

- **Population.** These are GitHub issue filers: mostly developers, reporting to a framework or sample maintainer. People who build with Lovable, Bolt.new or v0 and ask for help on Discord or Reddit are not here (those sources forbid automated collection or AI use).
- **Repository mix drives the numbers.** Azure has many threads because two Azure sample apps have busy trackers, not because Azure is harder. Compare hosts within a repo group, not across the whole file.
- **Search ranking.** GitHub's best-match ranking plus the per-repo cap decides what is in. It is not a random sample of all issues.
- **Resolution is a best guess.** `resolved` uses GitHub's `state_reason` and the author's own "it works now" messages. Many trackers auto-close stale issues; we try to detect it, but not always.
- **Language.** Mostly English; some Chinese (NextChat, Dify) and a few Portuguese and Spanish threads.
- **Thread types.** 66% of rows are problems, 13% feature requests, 12% questions, 4% docs requests, 4% other, under 1% CI. Filter on `thread_type == "problem"` for failure statistics.
- **Snapshot.** Counts and states are as of `collected_on`.

## Legal and ethics

- Collected through the GitHub REST API, within rate limits, read-only.
- GitHub's Acceptable Use Policy (section 7) lets researchers use public, non-personal information for research, provided resulting publications are open access. This dataset and its article are published openly. We store no personal information: no usernames, no names, no emails, no text.
- The issue text stays under its authors' rights on GitHub; we only link to it.
- **Removal:** if you filed one of these issues and want its row removed, open a discussion on this dataset page and we will remove it.

## License

Our coded data (the columns we produced) and the scripts: **CC BY 4.0**. The linked GitHub content is not included and is not covered by this license.

## Citation

Hal9 (2026). *AI App Cloud Deploy Issues (GitHub)*. Hugging Face dataset. The scripts, the coding rules and the sample review were made with AI help (Claude) in October 2026.