rob thijssen
3db1f586a8
Add newsfeed-modelwatch: a one-shot, systemd-timer-driven producer that watches
the Hugging Face Hub for open-weight releases and pushes filtered candidates to
the ingest endpoint. Answers "how do I automate the HF firehose into my feed" —
the RSS half already works on the pull rail; this is the JSON/API half.
How it works: each run polls the HF models API (a watchlist of orgs +
trendingScore), applies an admission predicate (license allowlist, total-params
cap, safetensors present, not gated, pipeline_tag), and POSTs the survivors to
POST /v1/ingest/candidates under a bearer token, tagged source="huggingface".
Two properties of the existing ingest rail shape the design:
- Ingest is idempotent on (user, external_id). Using the repo id as external_id
makes the producer STATELESS — no dedup table; re-runs/overlapping timers just
re-submit and the server drops repeats.
- HF `tags` are copied onto the candidate, so the user's per-interest weights do
the ranking. The predicate is only an ADMISSION filter (what's worth surfacing
at all) and stays subordinate to explicit weights, per the house rule.
Layout: newsfeed-modelwatch is a client of the API, not an internal component —
it holds no core/data deps. Pure logic (HF types, predicate, mapping to
CandidateSubmission, param formatting) lives in the lib with unit tests; the bin
does the HTTP polling/posting. CandidateSubmission gains Serialize so the
in-workspace producer can build and post one.
Ops:
- asset/systemd/newsfeed-modelwatch.{service,timer}: oneshot + 3-hourly timer.
The ingest token is a secret, kept in /etc/newsfeed/modelwatch.env
(NEWSFEED_TOKEN, mapped to `token` by figment) so a redeploy of the config
never clobbers it.
- asset/config/modelwatch.toml.tmpl: watchlist + predicate, no secret.
- infra-setup.sh installs the units, creates the env placeholder (never
overwritten), enables the timer, and prints the token step.
- deploy.yml builds + ships the binary and the (secret-free) config each deploy.
Verified: unit tests (gated union, license precedence, admit accept/reject,
mapping); dry-run against live HF; full loop against a local API — minted token,
submitted a real release, confirmed it landed in the feed with the huggingface
source linked and HF tags carried through. fmt/clippy/test all green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016fKZzDpvjiJ9eYbPGgJvUP
36 lines
1.4 KiB
Cheetah
36 lines
1.4 KiB
Cheetah
# newsfeed-modelwatch configuration (production). Rendered to /etc/newsfeed/modelwatch.toml
|
|
# by the deploy workflow.
|
|
#
|
|
# The ingest token is NOT here — it is a secret, supplied via NEWSFEED_TOKEN in
|
|
# /etc/newsfeed/modelwatch.env (created once by infra-setup.sh). That keeps redeploys of
|
|
# this file from ever clobbering the token.
|
|
|
|
ingest_url = "http://127.0.0.1:22672/v1/ingest/candidates"
|
|
source_name = "huggingface"
|
|
|
|
# HF orgs to watch for new open-weight repos.
|
|
authors = [
|
|
"tencent", "Qwen", "deepseek-ai", "mistralai", "zai-org", "moonshotai",
|
|
"openbmb", "allenai", "nvidia", "ibm-granite", "google", "meta-llama", "microsoft",
|
|
]
|
|
|
|
# The trending list is the best built-in noise filter; a notable release surfaces fast.
|
|
include_trending = true
|
|
trending_limit = 25
|
|
per_author_limit = 15
|
|
# Ignore repos older than this so the first run doesn't backfill old "newest" repos.
|
|
max_age_days = 30
|
|
|
|
http_timeout_secs = 20
|
|
user_agent = "newsfeed-modelwatch (+https://git.lair.cafe/grenade/newsfeed)"
|
|
|
|
# Admission filter: what's worth putting in front of you at all. Ranking is NOT decided
|
|
# here — the HF tags ride onto each candidate so your per-interest weights do the ranking.
|
|
[predicate]
|
|
licenses = ["apache-2.0", "mit"]
|
|
# Total safetensors params. NB: this is total, not MoE-active (that needs config.json math).
|
|
max_params = 40000000000
|
|
require_safetensors = true
|
|
skip_gated = true
|
|
pipeline_tags = ["text-generation"]
|