Skip to content

Add hf-free-maxxing skill: the verified Hugging Face free-tier operating manual (knowledge-only) - #271

Open
landogayatri wants to merge 3 commits into
huggingface:mainfrom
landogayatri:add-hf-free-maxxing-knowledge-only
Open

landogayatri wants to merge 3 commits into
huggingface:mainfrom
landogayatri:add-hf-free-maxxing-knowledge-only

Conversation

@landogayatri

Copy link
Copy Markdown

Re-submission of #264, restructured per @evalstate's feedback ("repo for hf skills, would not redirect to untrusted binaries") — this version is knowledge-only: a single SKILL.md, zero binaries, zero install steps. Every instruction is directly actionable with the official hf CLI, the official SDKs (huggingface_hub / @huggingface/hub / @huggingface/inference), or plain curl.

What it contributes

The live-verified free-tier operating layer that no existing skill covers:

  • The free-stack math: 100 GB private + 8.7 TB public storage per entity (orgs multiply pools), 20+30 TB/mo egress, $0.10/mo inference credits ≈ 5M tokens on the cheapest pinned lane, ZeroGPU 300 GPU-s + 8 runs/rolling-24h, unlimited static hosting, the POST /uploads media CDN that sits outside both storage and bandwidth quotas
  • 14 traps that cost money or hours if learned the hard way, e.g.:
    • create_bucket() defaults to public (verified live on huggingface_hub 1.9.2)
    • Router inference is price-blind unless you pin the provider (auto routing lands on paid lanes)
    • Placeholder-billing burst mechanics ($0.01 booked per request; ~10 in flight → transient 402s)
    • Credit reset semantics (calendar month, no rollover — verified live Oct 1, 2026)
    • ZeroGPU inputs must be base64 data-URIs; outputs are tmp capability-URLs
    • A README without sdk: static front-matter bricks a static Space
    • The security-checkup risk gate on credential surfaces
  • SDK-vs-raw routing: which surfaces the official SDKs cover (and the Python/JS asymmetry) vs which are raw-endpoint-only
  • Drift watch: promo lanes vanishing, pricing-schema migrations, silently-retired endpoints — with a monthly re-check recipe

Verification

Every number, endpoint, and failure mode was probed live on a free account (Sep–Oct 2026); evidence transcripts: https://huggingface.co/landogayatri/hf-free-maxxing. The curl recipes in the skill were re-tested against the live API this week (including the datasets-server examples).

No code ships in this PR — if a companion CLI is ever wanted, it can live outside this repo and be referenced separately.

…ing manual (knowledge-only)

Every free capability measured live (Sep-Oct 2026): storage math, the
$0.10/mo inference credits with cheapest-lane pinning, ZeroGPU quota
semantics, static hosting, the quota-free media CDN, dataset queries,
OIDC, registry, webhooks — plus the 14 traps that cost money or hours if
learned the hard way. Zero binaries: every instruction actionable via the
official hf CLI, official SDKs, or plain curl.
- SSE quick-win now sends the required `Accept: text/event-stream` header
  (plain curl gets a JSON error — live-verified)
- Prose rule 2: same header note
- Ternary placeholder trap: add the reversal semantics (~30 min, burns
  burst headroom while it sits)
- pip floor `>=1.9,<2.0` (matches what the evidence was gathered on) +
  optional boto3 note for SigV4 presign
- "S4" research jargon -> "Sep 2026"
… 8, 'Free Users: None', staff cited account farms) — banner + re-annotations (BYO/PRO framing, <=60s bucket-free harmonization, durable-mirror provenance). Unchanged at re-scan: storage, static hosting, CDN, ZeroGPU, data queries, identity.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant