Charts and quick start

Chart data and sources: assets/chart-data.json. No counts could be read from this card, so only the "not yet evaluated" status is drawn; the card's own evaluation text is unchanged.

Sizes are Hub file sizes on 2026-11-19. They say nothing about quality.
Quick start
from transformers import AutoTokenizer, AutoModelForCausalLMtok = AutoTokenizer.from_pretrained("SZLHOLDINGS/khipu-r3")model = AutoModelForCausalLM.from_pretrained("SZLHOLDINGS/khipu-r3", torch_dtype="auto")msgs = [{"role": "user", "content": "<controller-supplied context + request>"}]inputs = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt", return_dict=True)out = model.generate(**inputs, max_new_tokens=256, do_sample=False)print(tok.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
In transformers 5.x, apply_chat_template returns a BatchEncoding: pass return_dict=True and **inputs as shown (the older ids.shape[1] form raises AttributeError).
Run check (2026-11-19, transformers 5.23.0, torch 2.15.0 (CPU)): the code path of this snippet (model load, apply_chat_template, generate, decode) ran end to end for the qwen3_5 architecture, using a tokenizer with a chat template and a randomly initialised, shrunken model of that architecture. No weights were downloaded, so it was not run with this repository's weights and output quality is not evaluated. Private or gated repositories need hf auth login first.
Additional limitations
- Standard benchmarks (MMLU-Pro, GPQA, IFEval, BFCL): not yet evaluated.
- Safety evaluation (toxicity, jailbreak and red-team suites): not yet evaluated.
- This card's evaluation text was not charted because no count could be read from it unambiguously.
What this is
A separate laptop Blackwell research adapter based on Qwen/Qwen3.5-0.8B. The training_receipt.laptop-blackwell.json records bf16 LoRA with r=16 and alpha=32, 23 training rows from train.jsonl and train.abstain.jsonl, and train loss 0.4397. That loss is a training metric, not an evaluation.
Evidence and artifact identity
Reviewed Hub revision: 89807a2680e2400bbbc7523772aa6ff4b86a7988. The repository contains adapter_model.safetensors and model.safetensors; the historical training receipt's weights=LOCAL and push_to_hub=false describe that training step, not current file absence.
There is no eval_report.json in this Hub snapshot, and no committed Hub held-out receipt supporting the former README claim “grounding 4/5, abstain 0/6.” That unsupported claim is withdrawn.
The separate post-publication gate receipt names Hub revision 89807a2680e2400bbbc7523772aa6ff4b86a7988 and reports 1/5 on the canonical held-out suite in transformers-cpu mode, with overall FAIL, heldout_passed=false, productionAuthorization=false, and publicationAuthorization=false. This recorded canonical result does not substantiate the former 4/5 + 0/6 claim. It records refusalPassed=false and a refusal regression: rate 0.2 against baseline 0.4. The receipt retains runtime warnings, and earlier execution failures remain in the canonical gate archive. Therefore “no held-out evaluation exists anywhere” would also be inaccurate: the canonical source contains this later failed evaluation.
The merge_receipt.json reports a 2026-09-14 repair of a prior empty base-copy merge: 96/96 adapter modules applied, maximum weight delta 0.0013806568458676338. A nonzero merge delta does not establish behavioral quality.
What the evidence does and does not establish
The training_receipt.laptop-blackwell.json retains evals=none-this-run, publication_eligible=false, and autonomy_eligible=false. Its training-run adapter digest is receipt-reported and is not asserted as a direct hash of the currently published adapter file. Exact current artifact qualification does not follow from a training receipt, a merge receipt, or file presence. The later canonical gate names the reviewed Hub revision and returns FAIL. The provenance.json also retains blocked artifact-lineage, consent, privacy-review, deployment, and served-revision boundaries.
Release, autonomy, and deployment boundary
Not promotable. publication_eligible=false and autonomy_eligible=false remain unchanged. No production, deployment authorization, autonomous action, approval authority, or house-lab pin follows from these records. Lambda uniqueness remains Conjecture 1: open and advisory, never a theorem.
Relationship and source of truth
This SKU does not overwrite SZL-Khipu-1.5B, SZL-Khipu-1.5B-GGUF, KHIPU-R2, brain-navigator-r2, or SZL-Khipu-1.5B-BrainNavigator.
Use the immutable Hub training and merge receipts and the canonical failed evaluation receipt above. The historical source card and Khipu curriculum source provide context; their prose is not a substitute for an artifact-scoped evaluation receipt.
The estate
Provenance note: duplicate weights
The listed weight files in this repository are byte-identical (same LFS SHA-256; checkpoint-* excluded) to files in another repository: SZLHOLDINGS/khipu-r3-laptop-lora@8187bda1b27f (shared: adapter_model.safetensors) (checked 2026-11-12 at 8a5ccd68fb33). Identical bytes are one artifact published under more than one name; this repository is not independent evidence of a separate training run, and any results stated for it apply to that same artifact. Existing result lines above are left unchanged.
Limitations
- This model is a research candidate; it has not been deployed in production
- The estate's honesty battery (6 probes) governs promotion; see the receipts chain
- This model may fabricate on unanswerable questions; see the confabulation conjecture
- Intended for research use; not for safety-critical applications
Measured evidence: EVAL RECEIPTS BOUND (2026-11-14)
The estate published measurement receipts naming this repository in the dataset 'SZLHOLDINGS/model-eval-receipts' (dataset commit '44fd9dac4145a5e9e55f28df50390925d984631f'). Each is mirrored byte-identically under `receipts/eval/` with its sha256 in the table below. Estate doctrine: results publish verbatim, win or lose.
Table with columns: Receipt, schema, date, Measured headline, sha256| Receipt | schema | date | Measured headline | sha256 |
|---|
| `receipts/eval/khipu-r3-exact-bytes-eval-20261107.json` | szl.model-eval-receipt/v1 | 2026-11-07 | model=SZLHOLDINGS/khipu-r3; base=Qwen/Qwen3.5-0.8B; probes=6; abstained=1; answered=5; verdict=FAIL | `5fdc9a2eabf95429…` |
Verbatim honest notes (excerpted):
- khipu-r3-exact-bytes-eval-20261107.json: The exact-bytes evaluation of the khipu-r3 artifact FAILS the confabulation probes: 1/6 abstained, 5/6 answered, of which 4 fabricated estate-specific facts (invented benchmark certifier, invented contract license, fabricated paper author, fabricated book titl
- Binding scope: repository identity; mirrors are byte-identical, sha256-pinned to dataset commit 44fd9dac4145a5e9e55f28df50390925d984631f.
- These are measured results, not capability claims. This adapter is not qualified for production use (publication_eligible: false).
Measurement graph (receipt-bound, 2026-11-14)

- The chart above is generated from the byte-identical mirrored receipts in this repository (sha256-pinned in the SVG footer). It plots measured numbers verbatim, including failures - estate doctrine: win or lose, published as measured.
- publication_eligible: false. Not qualified for production use.
Curated repository facts (curated-card batch-3 2026-11-16)
- Observed Hub revision immediately before this edit:
472dd4c5df896ac851789e09f76e4df37cd7ad8c. The prior card remains in revision history.
- Observed repository creation date:
UNAVAILABLE.
- Observed repository contents: 19 files (2 weight files, 4 config/tokenizer files, 1 receipt files, 11 other files).
Receipts present:
receipts/eval/khipu-r3-exact-bytes-eval-20261107.json.
- License tag observed verbatim in this card's frontmatter:
apache-2.0. This edit changed no license and asserts none beyond the observed tag.
- Source mapping (estate asset registry): GitHub szl-holdings/szl-forge observed at HEAD
6c21496eaf346f8b2184b711335987c007a69ab0 (last push 2026-11-16T00:47:09Z); registry method card.
- Atelier classification: none recorded in the estate manifest. No class-based claim is made on this card.
- Eval battery receipt: none in this repository. No evaluation result, benchmark, or model-index metric is claimed in this edit.
- Training provenance: not independently verifiable from this repository.
- Hub reads at manifest generation (2026-11-14): 3.
- This edit adds observed repository facts only. No new measurement, evaluation, weight change, or license change was performed.