135 lines
4.7 KiB
Markdown
135 lines
4.7 KiB
Markdown
|
# Collection policy — open-weight reserve
|
|||
|
|
|
|||
|
|
**Status:** foundation
|
|||
|
|
**Related:** `schema.yaml`, `docs/backup-storage-policy.md`, `INTENT.md`
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Purpose
|
|||
|
|
|
|||
|
|
Decide **what** enters the open-weight reserve, **who** may approve it, and
|
|||
|
|
**when** a daily-brief candidate becomes a catalog entry with blobs on backup
|
|||
|
|
storage.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Goals
|
|||
|
|
|
|||
|
|
* Keep a **small, high-leverage** reserve — not a Hugging Face mirror
|
|||
|
|
* Enforce **license and integrity** before download completes
|
|||
|
|
* Match capacity to the backup storage soft quota
|
|||
|
|
* Prefer models that serve axes **B** and **C**, plus strategic **A** open releases
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Eligibility (must pass all)
|
|||
|
|
|
|||
|
|
1. **Open weights** — weights obtainable under terms that allow offline retention for lab use
|
|||
|
|
2. **Clear license** — SPDX or linkable license text; `allows_offline_retention: true`
|
|||
|
|
3. **Stable provenance** — official org, tagged release, or commit revision (not anonymous drive-by reupload as sole source)
|
|||
|
|
4. **Lab rationale** — written `reason` tied to at least one axis A–D (usually B/C)
|
|||
|
|
5. **Capacity** — estimated size fits under remaining soft quota (see backup storage policy)
|
|||
|
|
|
|||
|
|
Fail any gate → status `rejected` with reason, or never enter catalog.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Priority rubric
|
|||
|
|
|
|||
|
|
| Priority | Guidance |
|
|||
|
|
| -------- | -------- |
|
|||
|
|
| **high** | Rare or strategically important; license/access risk of disappearance; uniquely strong for B/C at our hardware class; hard to re-obtain |
|
|||
|
|
| **medium** | Clear lab use within 1–2 quarters; good quality/cost; easy enough to re-download but worth having cold |
|
|||
|
|
| **low** | Nice to have; only collect if quota headroom is large and pull is cheap |
|
|||
|
|
|
|||
|
|
Daily brief **collection candidates** should set a suggested priority; approval may change it.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Approval rule of thumb
|
|||
|
|
|
|||
|
|
| Estimated total size | Approval |
|
|||
|
|
| -------------------- | -------- |
|
|||
|
|
| **< 5 GiB** | Operator or lab agent may collect after license check; catalog entry required before or immediately after |
|
|||
|
|
| **5–40 GiB** | Explicit operator approval (chat, workplan task, or signed catalog `approved_by`) |
|
|||
|
|
| **> 40 GiB** | Operator approval **plus** check against soft quota and whether a smaller quant/variant suffices |
|
|||
|
|
| **Any size if quota ≥ 70% used** | Operator approval required regardless of size |
|
|||
|
|
| **Unclear license or ToS risk** | Do not collect; status `rejected` |
|
|||
|
|
|
|||
|
|
“Operator” means the human lab owner (or a documented delegate). Agents may
|
|||
|
|
**nominate** (`status: candidate`) freely from briefs; they may **collect** only
|
|||
|
|
within the < 5 GiB band when licenses are unambiguous — otherwise stop at
|
|||
|
|
`candidate` / `approved`.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Lifecycle
|
|||
|
|
|
|||
|
|
```text
|
|||
|
|
brief nominates
|
|||
|
|
→ candidate (catalog YAML, no blobs required)
|
|||
|
|
→ approved (license + size + quota OK)
|
|||
|
|
→ collecting (download in staging/)
|
|||
|
|
→ collected (blobs complete, checksums recorded, storage_path set)
|
|||
|
|
→ verified (optional re-hash / smoke load)
|
|||
|
|
→ superseded|evicted (replaced or removed; metadata kept)
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
Rejected candidates stay in catalog only if useful as a decision record; otherwise omit.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## What we prefer to collect
|
|||
|
|
|
|||
|
|
* Small/mid instruct and code models that fit the hardware envelope
|
|||
|
|
* Strong embedding / rerank models for local RAG
|
|||
|
|
* Base models known to fine-tune well under QLoRA/LoRA on lab GPUs
|
|||
|
|
* Official quant releases when they are the supported distribution
|
|||
|
|
* Adapters and tokenizers that unlock a reserved base (as companions)
|
|||
|
|
|
|||
|
|
## What we usually skip
|
|||
|
|
|
|||
|
|
* Duplicate quants of the same revision already reserved
|
|||
|
|
* Huge models with no near-term local run/train path and no access-risk story
|
|||
|
|
* Merges/repacks without provenance
|
|||
|
|
* Datasets larger than model weights unless separately justified (default: out of band)
|
|||
|
|
* Anything requiring acceptance flows we cannot satisfy offline
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Companions
|
|||
|
|
|
|||
|
|
Tokenizers, LoRA adapters, and small eval fixtures may be collected when:
|
|||
|
|
|
|||
|
|
* they are required to use a reserved base, or
|
|||
|
|
* they are small (< 1 GiB) and high leverage
|
|||
|
|
|
|||
|
|
Link via `companions` in the catalog schema.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Brief integration
|
|||
|
|
|
|||
|
|
1. Brief section **Collection candidates** nominates items.
|
|||
|
|
2. Operator/agent opens `inventory/catalog/{id}.yaml` with `status: candidate`.
|
|||
|
|
3. Approval and download follow this policy and `docs/backup-storage-policy.md`.
|
|||
|
|
4. Brief `brief_refs` on the entry point back to the nominating day(s).
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Eviction rule of thumb
|
|||
|
|
|
|||
|
|
When over quota or cleaning:
|
|||
|
|
|
|||
|
|
1. `low` priority, easily re-obtainable from still-live official URLs
|
|||
|
|
2. Superseded revisions with a newer `verified` replacement
|
|||
|
|
3. Never silent-delete: set `status: evicted`, clear or note `storage_path`, append `history`
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Non-goals
|
|||
|
|
|
|||
|
|
* Automatic bulk mirrors of entire orgs
|
|||
|
|
* Collecting on every brief mention without priority
|
|||
|
|
* Bypassing license gates for “research only” convenience
|