Update dataset card

README.md

@@ -1,6 +1,6 @@
 ---
 license: mit
-pretty_name: Verified free proxies – daily snapshots
+pretty_name: Checked free proxies – daily snapshots
 tags:
 - proxies
 - networking
@@ -15,32 +15,40 @@ configs:
   data_files: "data/*/*.parquet"
 ---
 
-# Verified free proxies – daily snapshots
+# Checked free proxies – daily snapshots
 
 One Parquet file per UTC day with every public HTTP, SOCKS4 and SOCKS5 proxy that passed all checks in the
-first hourly run of that day. Collected by [proxy-scraper](https://github.com/maximilianfeix/proxy-scraper),
-which pulls from 700+ public lists and keeps only proxies that actually relay traffic.
+first published run of that day. Collected by [proxy-scraper](https://github.com/maximilianfeix/proxy-scraper),
+which pulls from hundreds of public lists and keeps only the entries that pass every check. All rows are entries
+that passed: candidates that failed or were not checked are not in the data. An entry is a `type://ip:port`
+combination, not a machine – many are ports on the same host.
 
 ```python
 from datasets import load_dataset
 
 ds = load_dataset("Taventix/proxy-scraper-snapshots", split="train")
 df = ds.to_pandas()
-# residential HTTPS exits that got through to Reddit, by country
-residential = df[(df.hosting == False) & df.works_on.apply(lambda sites: "reddit" in sites)]
-residential.groupby("country").size().sort_values(ascending=False).head()
+# non-hoster exits that got through to Reddit, by country
+non_hoster = df[(df.hosting == False) & df.works_on.apply(lambda sites: "reddit" in sites)]
+non_hoster.groupby("country").size().sort_values(ascending=False).head()
 ```
 
 ## How a proxy gets in
 
-Each hour about 150,000 candidates are checked, and a proxy is kept only if it
+About every hour 150,000 candidates are checked, out of roughly 1.2–1.4 million listed protocol/address
+combinations. Proxies that passed before and sources with a good hit rate go first. An entry is kept only if it
 
-1. completes a real protocol handshake (HTTP CONNECT, SOCKS4 or SOCKS5),
-2. loads two unrelated sites with the same exit IP – scanner-only honeypots fail here,
-3. returns a known page byte for byte – proxies that inject scripts or ads fail here,
-4. for `https = true`, tunnels TLS with a certificate that verifies.
+1. completes a real protocol handshake (HTTP CONNECT, SOCKS4 or SOCKS5) and a first request that returns a public
+   exit IP that is not ours,
+2. shows the same exit IP on a second request to an unrelated site. A failure means only that the second request
+   did not confirm the proxy: an unstable proxy, an exit that changes between requests, a proxy that only answers
+   the first check, or the second site refusing it,
+3. returns one static test page byte for byte over plain HTTP. A failure means the bytes differed; why is not
+   determined,
+4. for `https = true`, tunnels to one test host on port 443 with a TLS certificate that verifies, at check time.
 
-Of 3.6 million checks in one day, 1.1 % passed. A day's snapshot holds between 1,000 and 2,200 proxies.
+Of 3.6 million checks of that prioritised sample in one day, 1.1 % passed. A day's snapshot holds between 1,000
+and 2,200 proxies.
 
 ## Fields
 
@@ -53,18 +61,18 @@ Of 3.6 million checks in one day, 1.1 % passed. A day's snapshot holds between 1
 | `country` | string | Two-letter country code of the exit IP |
 | `asn` | int64 | Autonomous system number of the exit IP |
 | `org` | string | Network / provider that owns the exit IP |
-| `hosting` | bool | Exit IP is in a datacenter or hosting network |
+| `hosting` | bool | The provider name of the exit IP looks like a hosting company (datacenter) |
 | `blocklisted` | bool | Exit IP is on the SpamCop blocklist |
 | `anonymity` | string | `elite` (target sees no proxy), `anonymous` or `transparent` |
-| `https` | bool | Tunnels HTTPS with a verified TLS handshake |
+| `https` | bool | Passed the HTTPS test: a tunnel to one test host with a TLS certificate that verifies, at check time |
 | `latency` | int64 | Response time in the check, milliseconds |
 | `speed_kbps` | int64 | Download speed in KiB/s (HTTPS-capable proxies only, measured since 2026-09-28) |
 | `works_on` | list of strings | Big sites it got through to in that run: google, reddit, amazon, instagram, tiktok, discord |
 | `blocked_on` | list of strings | Big sites that turned it away in that run |
-| `streak` | int64 | Consecutive hourly runs it has been on the list |
-| `uptime_24h` | int64 | Share of the last 24 hourly runs it passed, percent |
-| `uptime_7d` | int64 | Share of the last 7 days' hourly runs it passed, percent |
-| `first_seen` | string | When it first passed, ISO 8601 UTC |
+| `streak` | int64 | Consecutive published runs (about hourly) it has been on the list |
+| `uptime_24h` | int64 | Share of the published runs of the last 24 hours in which it was on the list, percent |
+| `uptime_7d` | int64 | Share of the published runs of the last 7 days in which it was on the list, percent |
+| `first_seen` | string | When it was first on the list, ISO 8601 UTC |
 | `snapshot_date` | string | The UTC day of the snapshot, `YYYY-MM-DD` |
 
 Sites are only checked for proxies with `https = true`, so both lists are empty for the others. For an
@@ -73,22 +81,27 @@ amazon from the start, instagram and tiktok since 2026-09-28, discord since 2026
 measured yet on early days are null, and so are `asn`, `org`, `country`, `anonymity` and `exit_ip` when they're unknown. Every file has
 the same columns, so all days load as one table.
 
+Not being on the list in a run is not a measured failure, so `streak` and `uptime_*` describe list membership.
+From proxy-scraper 1.30 on, each row of the published `proxies.json` and of the daily `proxies-YYYY-MM-DD.json.gz`
+snapshots in the GitHub release also carries `checked_at` and `checks`: how each check ended (`passed`, `failed`,
+`unavailable` or `skipped`). The Parquet files here keep the fixed columns above.
+
 ## What it's good for
 
 - **Security monitoring:** open proxies that relay traffic right now – search them in login and outbound logs.
-- **Churn and lifetime:** most free proxies are gone within an hour; `streak`, `uptime_*` and `first_seen` show
-  how long they last.
+- **Churn and lifetime:** `streak`, `uptime_*` and `first_seen` show how long entries stay on the list. Most of
+  the entries that leave it had been on it for a single run.
 - **Who hosts open proxies:** `asn`, `org`, `hosting` and `country` over time.
-- **Site blocking:** how Google, Reddit, Amazon, Instagram, TikTok and Discord treat datacenter vs. residential exits.
+- **Site blocking:** how Google, Reddit, Amazon, Instagram, TikTok and Discord treat datacenter exits vs. other exits.
 
 ## Caveats
 
 - Free proxies are run by strangers: misconfigured servers, infected machines and traps. Don't send credentials
   or personal data through them.
 - Checks run from GitHub Actions in the US; a proxy can behave differently from another network.
-- A snapshot is one hourly run per day; the live list on the
+- A snapshot is the first published run of each UTC day; the live list on the
   [`proxy-list` branch](https://github.com/maximilianfeix/proxy-scraper/tree/proxy-list) and in
-  [free-proxy-list](https://github.com/maximilianfeix/free-proxy-list) is rebuilt every hour.
+  [free-proxy-list](https://github.com/maximilianfeix/free-proxy-list) is rebuilt about every hour.
 
 This card is kept in the [proxy-scraper repository](https://github.com/maximilianfeix/proxy-scraper/blob/main/docs/huggingface/README.md)
 and uploaded with each daily snapshot. MIT licensed.