KHIPU-R2 / README.md
betterwithage's picture
docs: honest ROADMAP silhouette
856a958 verified
|
Raw
History Blame
2.77 kB
metadata
license: apache-2.0
tags:
  - roadmap
  - not-a-checkpoint
  - szl-holdings
  - series-a
  - doctrine-v11
  - khipu

KHIPU-R2

ROADMAP. This Hub ID is the named next Khipu revision. There are no weights on this ID. Do not from_pretrained("SZLHOLDINGS/KHIPU-R2"). Live weights are SZL-Khipu-1.5B, not this ID. This repo does not overwrite the signed 1.5B.

Train scripts may exist here. Scripts are not a checkpoint. Training is ROADMAP. Weights are UNAVAILABLE.

What

  • A reserved name for a future Khipu revision (abstain retrain / successor line).
  • Files on Hub 2026-08-28 1:18 PM ET: train_khipu_r2.py, train_khipu_abstain.py — source, not tensors.

What-NOT

  • Not a trained next Khipu. Not an adapter on Hub. Not base_model: Qwen/Qwen2.5-1.5B-Instruct with relation adapter.
  • Not a replacement for SZL-Khipu-1.5B. Never overwrite that ID from this one.
  • Not MEASURED capability. A job id is not a checkpoint.

Status

Thing Label Method / N / date / what-NOT
Weights UNAVAILABLE Hub file list 2026-08-28 1:18 PM ET. N=0 safetensors / GGUF / adapter. Files present: README + two train scripts.
Training ROADMAP Scripts REACHABLE on this repo. No signed training_receipt + tensors together on Hub. What-NOT: a script is not a trained model.
HF Job REPORTED Job id 6a91bf11984507d9db4ea104 (link) — REPORTED from the prior card / Hub jobs, 2026-08-28. What-NOT: REPORTED is not MEASURED; this does not prove adapters exist on this ID.
Evals ROADMAP No eval receipt on this ID. N=0. Do not reuse Khipu 1.5B k/n as R2 scores.
Inference UNAVAILABLE Lab load forbidden. Lab serves signed GGUF only.

Live weights (different IDs)

Tensors that actually load today:

Doctrine

Doctrine v11. Λ (lambda) = Conjecture 1 — advisory, never green, never a theorem.

Honesty labels: MEASURED · REPORTED · DERIVED · MODELED · SNAPSHOT · REACHABLE · UNAVAILABLE · ROADMAP · CONJECTURE. Every number carries method / N / date / what-NOT. No model-index. No invented evals.