Artifact identity (noted 2026-09-30). Besides the LoRA adapter, this repository's root carries a merged full-precision checkpoint (
model.safetensors+config.json, loadable withtransformers) produced by the receipted CPU merge of this adapter into its declared base (merge_receipt.json). The adapter remains the artifact of record for every figure on this card; the merged bytes carry no separate held-out receipt and add no claim. Metadata-only note.
KHIPU-R2
Research-only · historical abstention 3/6 · publication and autonomy gates closed
Adapters are on this repo. Abstain is MEASURED 3/6, not a pass.
That statement describes the dated owner-run records below, rather than
a new evaluation or current model qualification. This is a separate
abstention-retraining SKU. Not a replacement for SZL-Khipu-1.5B.
Card and metadata review snapshot: 9fc2f7bebeaf4b3d6d26f7a2a1788a27f8c9ed52. No weights were downloaded, no model was loaded, and no evaluation or signature verification was rerun in this review.
Distributed forms
| Form | Files | Scope |
|---|---|---|
| QLoRA adapter | adapter_model.safetensors, adapter_config.json |
The PEFT example below explicitly selects this adapter form. |
| Separate fp32 merged checkpoint | Root model.safetensors, config.json, merge_receipt.json |
A distinct runtime form; selecting the adapter does not load this merged checkpoint. |
Current artifact bytes and their equality to historical evaluated bytes
were not independently rehashed. The merge receipt's base revision is
main@merge-time, which is not an immutable base identity.
Dated evaluation records
The 2026-08-28 owner-run job record
reports in-process Unsloth generation at temperature 0 with scoring ported
from eval_khipu.py. It reports held_out_in_gradients: false.
| Metric | Historical result | Limit |
|---|---|---|
| plan-valid | 11 / 11 | Small owner-run synthetic protocol |
grounding (eval.jsonl navigate) |
5 / 5 | Five navigation cases |
abstain (adversarial.jsonl) |
3 / 6 | Visible release blocker; not a pass |
| hallucinated citations | 0 | These cases only |
The 2026-09-13 corrected-harness CPU re-verification reports 11/11 plan-valid, 5/5 grounding, and 3/6 abstention on rebuilt fp16 merges and the published fp32 merge. The owner fixed a decoding harness that returned prompt characters instead of generated output; all prior fp16 scores produced by that buggy harness are void. The corrected runs are separate owner records and do not promote the model.
No signed R2 eval receipt in this atelier. The linked owner-run records are not an independent evaluation or public leaderboard. GPU UNAVAILABLE describes the CPU-only re-verification record and the earlier card state; no current hardware availability was tested. Publication and autonomy eligibility remain false.
The original Khipu 2/6 abstention record and R2's 3/6 record remain separate. No broad quality score or general safety guarantee follows.
Historical training
The prior card and training receipt
describe QLoRA r=32, α=64, seed 11, learning rate 2e-4 and 45 epochs,
on 15 navigation rows plus 8 abstention rows oversampled four times.
The reported train loss 0.017188… is a training metric, not evaluation.
The historical adapter digest is
e44d53f29f2d443598e06d6c0441557fd3a5010888c7aa97b56ec3c0e050d349;
it was not recomputed here. The earlier card records successful job
6a91bf11984507d9db4ea104 and prior error 6a91ba2c45686a1580c12020
(Trackio 404). This review did not query those jobs or rerun training.
Adapter usage
This example is untested in this review. An owner-verified immutable base revision is required; none is invented here. The blank base revision fails before imports or remote loading. The adapter repository is pinned to the reviewed documentation and metadata snapshot; this is not an evaluation binding or release approval.
import re
BASE_REVISION = "" # Supply an owner-verified immutable base commit.
if not re.fullmatch(r"[0-9a-f]{40}", BASE_REVISION):
raise ValueError("An owner-verified immutable base revision is required")
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base_id = "Qwen/Qwen2.5-1.5B-Instruct"
adapter_id = "SZLHOLDINGS/KHIPU-R2"
adapter_revision = "9fc2f7bebeaf4b3d6d26f7a2a1788a27f8c9ed52"
tok = AutoTokenizer.from_pretrained(base_id, revision=BASE_REVISION)
base = AutoModelForCausalLM.from_pretrained(base_id, revision=BASE_REVISION)
model = PeftModel.from_pretrained(base, adapter_id, revision=adapter_revision)
Outputs are proposals requiring external validation and approval. Lab load remains forbidden. A11OY-MINI contains separately scoped legacy and R2 evidence and inherits no KHIPU-R2 score or qualification.
Apache-2.0 is declared in metadata. No standalone LICENSE is listed at
the reviewed revision; artifact license coverage was not independently
verified. A license addition requires source-owner review of that coverage.
Canonical card authoring source: khipu-r2/card/README.md. Doctrine v11 LOCKED 749/14/163. Λ = Conjecture 1, advisory, never a theorem.
- Downloads last month
- 126
from peft import PeftModel from transformers import AutoModelForCausalLM base_model = AutoModelForCausalLM.from_pretrained("unsloth/Qwen2.5-1.5B-Instruct-bnb-4bit") model = PeftModel.from_pretrained(base_model, "SZLHOLDINGS/KHIPU-R2")