Hugging Face's logo

SZLHOLDINGS
/
KHIPU-R2

Text Generation
PEFT
Safetensors
English
qwen2
qlora
governed-agent
proposal-only
research-only
szl-holdings
khipu
abstain-retrain
conversational
Model card Files
xet
 Community
2
KHIPU-R2 / README.md
betterwithage's picture
betterwithage
docs: publish KHIPU-R2 card from szl-forge@55f56487a52f3f3cf6f32c0e5ee3637660dfc90a
1c48e77 verified
preview code
|
Raw History Blame Contribute Delete
11.7 kB
metadata 
thumbnail: >-
  https://huggingface.co/SZLHOLDINGS/KHIPU-R2/resolve/9fc2f7bebeaf4b3d6d26f7a2a1788a27f8c9ed52/og-card.png
license: apache-2.0
language:
  - en
base_model: Qwen/Qwen2.5-1.5B-Instruct
base_model_relation: adapter
library_name: peft
pipeline_tag: text-generation
tags:
  - qlora
  - peft
  - governed-agent
  - proposal-only
  - research-only
  - szl-holdings
  - khipu
  - abstain-retrain
szl:
  doctrine: v11-LOCKED
  lean: 749/14/163
  lambda: Conjecture 1 — advisory, never a theorem
  artifact_class: ADAPTER
  publication_eligible: false
  autonomy_eligible: false
  never_overwrite: SZLHOLDINGS/SZL-Khipu-1.5B
  jobs: COMPLETED
  job_id: 6a91bf11984507d9db4ea104
  job_prior_error: 6a91ba2c45686a1580c12020
  weights: AVAILABLE
  evals: MEASURED
  gpu: UNAVAILABLE

SZL Holdings

KHIPU-R2

A separate abstention-retraining experiment with adapter-scoped historical evaluation records.

Artifact: QLoRA adapter; separate merged checkpoint · Stage: Research · release blocked

Command Lab · Canonical source · Reviewed evidence

Provider compatibility support requested on 4 October 2026. No provider was listed in the dated Hub observation; adoption and use qualification remain unchanged.

Before you use it

  • Recorded abstention is 3/6, not a passing gate. Publication and autonomy eligibility remain false.
  • The adapter is the artifact of record for the figures; the merged checkpoint has no separate held-out qualification.
  • The recorded merge-time base reference is not an immutable identity. This experiment does not replace Khipu 1.5B.

Evidence passport

Dated input observation: 2026-11-13T22:16:36.754444+00:00. Existing research limits remain in force.

Inspect the immutable repository snapshot.

Observation Scope
Repository revision 3a698a0f9a96eb80c38b11a77e114f318f6e4820
Historical README retrieval HTTP 200; not a readback of this proposed panel
Recognized artifact filenames 2 filename hints; not tensor verification
Structured records retrieved 4; bounded selection, not an exhaustive evidence census

Historical input README SHA-256: 641ef554d5261d4655db87ce194c5139ebfa89e0aafc20bd0d63a16e9d97b99d.

Reviewed authoring baseline. Existing publisher: publish-khipu-card.yml. Source changes still require normal review and exact deployed-byte readback.

Inspect the records

Record Retrieved-byte SHA-256
training_receipt.json cd3b5f8b8f083314abd3bc3eee46aa2a6be8491571aa867860e5b4f85c4ed6ce
eval_measured.json e573dfba96dafd8ab0baa65a41bcf6878488d7ea2677d674c0b702b2878b5371
eval_verified_2026-09-13.json 122a19192ef72868d91ca3bbc59a460626dcf64d951e9c2ad2633d7f0f6b38a0
merge_receipt.json cc1fe8b33e835fef22bea6605386b3135185ea223453caf0c6cf416c95340511

Interpretation

Documentation and retrieved-record integrity only. No new training, model evaluation, weight-byte verification, independent signer identity, runtime qualification or promotion. Missing evidence is not a zero score. Training loss is not held-out performance. A copied receipt does not qualify a quantization or merged checkpoint. Existing negative results, license and technical disclosures remain authoritative.

Technical details and evidence

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 with transformers) 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.

SZL Orbit artwork: source · CC BY 4.0. Software and artifact licensing remain declared above.

Publication restoration note (2026-11-14)

On 2026-11-14, an off-route automated job stamped the public SZLHOLDINGS Hub model cards with a marketing template that replaced curated card bytes, including per-model claims this artifact has not earned. This source card in szl-holdings/szl-forge is the canonical reviewed card; that job did not alter it. This note marks the card so the closed-profile publisher republishes the reviewed bytes, banner, and source binding to the Hub with exact readback. The incident and the per-model evidence manifest are documented in szl-forge under docs/model-card-overhaul-20261114/. No qualification, evaluation, weight, adapter, config, visibility, hardware, or runtime state changed by this note.