quge007/ai-infrastructure-problem-observatory

AI Infrastructure Problem Observatory: Topics Dataset

Quick Summary

80,257 real-world AI infrastructure problems extracted from GitHub discussions, technical forums, and industry blog posts. Covers 22 months (2024-12 to 2026-09). Fully self-contained with 100% original source data embedded—no external joins needed.

Use this dataset to:

• Understand production challenges faced by AI teams (infrastructure, deployment, optimization)
• Identify emerging problems and their temporal trends
• Build problem-aware systems (chatbots, documentation, product features)
• Research AI infrastructure needs and solution adoption patterns

Dataset Overview

Metric	Value
Total Topics	80,257
Time Period	2024-12 to 2026-09 (22 months)
Size	462 MB (Apache Parquet, zstd compressed)
Sources	3 types (GitHub, forums, blogs)
Data Preservation	100% (complete original records embedded)

Data Breakdown

Source	Topics	Examples
GitHub	71,742	Issues/discussions from vllm, dify, litellm, sglang, ragflow, ollama, langfuse, openhands, tensorrt, lmcache, langgraph, dynamo
Technical Forums	7,357	HuggingFace Discussions, Ray Discourse, Kubernetes Discuss
Industry Blogs	1,154	Official posts from Anthropic (23), Manus (117), Together (149)

What's In This Dataset?

Each row represents one extracted problem topic with:

Core Content

• topic_text — Complete, coherent problem statement (full text reconstructed from source passages)
• topic_title — Concise problem title (complete sentence, no fragments)
• source_id — Originating source (e.g., "github_vllm", "forum_ray", "blog_anthropic")
• source_record_id — Original record ID (issue #, post ID, URL)
• created_date_utc — When the problem was reported

Full Source Preservation

• raw_source_record_json — Complete original record (JSON-encoded):
  • For GitHub/forums: Full API response (all fields, comments, metadata)
  • For blogs: Complete article with raw HTML body
  • No data loss: Use this field for analysis requiring original context

Traceability

• source_block_ids — Which passages from source formed this topic
• source_blocks_raw — The exact quoted passages (JSON-encoded)

Metadata

• team_name — Inferred company/team (from URL or context; may be incomplete)
• source_language — Language code
• source_url — Direct link to original source

Full schema: See schemas/topics_schema.json (52 columns total)

Why 100% Original Data?

This dataset includes complete source records because:

• Extracted topics may lose context; the original source is the source of truth
• Researchers often need the full conversation, not just the problem statement
• Complete data enables reproducibility and custom re-analysis
• Supports fact-checking and context verification

Result: Each topic is independently publishable — no need to reference original sources separately.

How to Use

Load a Single Month

import pyarrow.parquet as pq

# Load September 2026 data
table = pq.read_table("database/topics/2026-09/part-00000.parquet")
print(f"Records in 2026-09: {table.num_rows}")

# Convert to pandas for exploration
df = table.to_pandas()
print(df[["topic_title", "source_id", "created_date_utc"]].head(10))

Access Complete Original Data

import json

# Get the full original record for any topic
row = df.iloc[0]
original_source = json.loads(row["raw_source_record_json"])

# For GitHub issues: access all original fields
if row["source_id"].startswith("github"):
    print(f"Issue title: {original_source.get('title')}")
    print(f"Author: {original_source.get('user', {}).get('login')}")
    print(f"Comments count: {original_source.get('comments')}")

# For blog articles: raw HTML body
if row["source_id"].startswith("blog"):
    html_body = original_source.get("body")
    print(f"Article length: {len(html_body)} chars")

Filter by Source Type

# GitHub issues/discussions only
github_topics = df[df["source_id"].str.contains("github", case=False)]
print(f"GitHub topics: {len(github_topics)}")

# Forums only
forum_topics = df[df["source_id"].str.contains("hugging|ray|kubernetes", case=False)]
print(f"Forum topics: {len(forum_topics)}")

# Blog articles only
blog_topics = df[df["source_id"].str.contains("blog_", case=False)]
print(f"Blog topics: {len(blog_topics)}")

Temporal Analysis

# Count topics by month
monthly_counts = df.groupby(df["created_date_utc"].dt.to_period("M")).size()
print(monthly_counts)

# Find problems reported in specific time ranges
recent = df[df["created_date_utc"] > "2026-06-01"]
print(f"Topics from June 2026 onwards: {len(recent)}")

Data Quality & Completeness

✅ Verified completeness:

• GitHub/Forum: All records include original API JSON (100% fields preserved, 0 null values in raw_source_record_json)
• Blog articles: 289/289 articles successfully extracted and consolidated with raw HTML body
• No filtering: All extracted topics included; no row-level filtering or sampling

⚠️ Known characteristics:

• Blog extraction: Uses LLM-based extraction; potential for semantic interpretation variance (see schema for confidence indicators if available)
• Multi-topic articles: Blog articles may yield multiple topics; this is intentional (preserves distinct problems in single article)
• Temporal coverage: Fixed window (2024-12 to 2026-09); no real-time updates

Data Collection Methodology

Collection Window: 2025-01-01 to 2026-09-16 (UTC)

GitHub Issues & Discussions

• Automated API collection from 12 key AI infrastructure repositories
• All public issues and discussions during window
• Repositories: vllm, dify, litellm, sglang, ragflow, ollama, langfuse, openhands, tensorrt, lmcache, langgraph, dynamo

Technical Forums

• HuggingFace Discussions, Ray Discourse, Kubernetes Discuss
• All public forum topics during window

Industry Blog Posts

• Manual curation: 289 official technical blog articles
• Sources: Anthropic, Manus, Together
• Covers product updates, architecture decisions, lessons learned

Topic Extraction

• GitHub/Forums: Topics already at problem grain (one issue/post → one topic)
• Blog articles: LLM-assisted extraction of distinct problems, consolidated by title
• Result: 80,257 consolidated topics with complete provenance

Limitations & Caveats

1. Language: Majority English-language; other languages underrepresented
2. Deduplication: Same problem discussed in multiple sources appears as separate rows (preserves frequency signal; not semantically deduplicated)
3. Temporal: Window is frozen (2026-09-16); no real-time updates
4. Team attribution: team_name field inferred from URL/context; may be incomplete or inaccurate
5. Blog content: LLM extraction may introduce minor semantic variation vs. original text

Citation

If you use this dataset, please cite:

@dataset{ltp2026_ipo_topics,
  title = {AI Infrastructure Problem Observatory: Topics Dataset},
  author = {{LTP Research Team}},
  year = {2026},
  url = {https://huggingface.co/datasets/quge007/industry-problem-observatory},
  license = {CC-BY-4.0}
}

Or use the provided CITATION.cff file for other formats.

License

This dataset is released under Creative Commons Attribution 4.0 International (CC-BY-4.0).

You are free to:

• Share and use this dataset for any purpose (commercial, research, education)
• Create derived works and adaptations
• Combine with other datasets

You must:

• Give appropriate credit to the dataset creators
• Link to the license
• Indicate if modifications were made

See LICENSE file for full legal text.

Getting Help

Common Questions

Q: How do I use this with my AI project? A: Load one month at a time using PyArrow, filter by source type, and extract the raw_source_record_json field for original context. See "How to Use" section above.

Q: Can I redistribute this dataset? A: Yes, under CC-BY-4.0. You must attribute the original creators and link to the license.

Q: Why are some problems duplicated across rows? A: Intentional. If one problem is discussed in both GitHub and a blog post, it appears twice. This preserves frequency information and multiple perspectives on the same issue.

Q: How do I understand the data fields? A: See schemas/topics_schema.json for complete field definitions and types.

Q: Where do I find recent topics? A: Most recent data is in topics/2026-09/part-00000.parquet (September 2026). Load and explore by created_date_utc.

---

Last Updated: 2026-09-30
Data Version: 1.0.0
Schema Version: topics.v2