Modelpublic

SZLHOLDINGS/szl-receiptagent-qwen35-0.8b-v2

sourceHugging Faceapache-2.0updated 11d agoView on Hugging Face

1likes470downloads

Model Card

<p align="center"><em>A receipt agent you can retrain between coffee and lunch.</em></p>

SZL ReceiptAgent Qwen3.5 0.8B v2 is a small, proposal-only adapter for drafting structured governance receipts and refusing requests that would fabricate evidence, approval, execution, or measured values.

The cut

Everyone fine-tunes 8B because it looks serious. We fine-tune 0.8B because the receipt loop should be cheap enough to run every night.

A receipt agent you can retrain between coffee and lunch, with the same doctrine as the 1.5B.

Silhouette → leave → SZL

Leader	Take, then tweak
Anthropic	Same constitution, smaller body.
NVIDIA	Small NIM-shaped thing, without NIM.
Unsloth	This is the Unsloth poster: FastLanguageModel, Qwen3.5-0.8B, PEFT.

Nobody else ships this combination. That is the point of a one-of-one.

Intended use

Nightly receipted Unsloth loop.

Limitations

—Adapter only. Load on the disclosed base.
—No signed eval in this atelier.

Canonical GitHub: `szl-holdings/szl-forge`

Intended use

Use this adapter behind a validating controller that:

1.validates every draft against the published JSON schema;
2.requires policy and human approval outside the model;
3.executes actions outside the weights; and
4.mints a cryptographic receipt only after approved execution.

The adapter is not an autonomous agent, authorizer, executor, factual oracle, or substitute for source retrieval.

Exact lineage

—Canonical base: Qwen/Qwen3.5-0.8B
—Base revision: 2fc06364715b967f1860aea9cf38778875588b17
—Training implementation: unsloth/Qwen3.5-0.8B
—Implementation revision: 23c69c53358a07516b5827588b3fdb12ae78fd65
—Runtime: Unsloth FastVisionModel
—License: Apache-2.0

Training

—Hardware: NVIDIA GeForce RTX 5050 Laptop GPU
—Optimizer steps: 64
—Admitted training rows: 37
—Trainable parameters: 10,822,656
—Total parameters observed: 863,808,576
—Final aggregate training loss: 0.9142942871840205
—Peak reserved training memory: 1,671,430,144 bytes
—Training report SHA-256: 1d1ce062e76aeccabe75fabb2c74d8bdddfc6f9a86f13c5cfd9960e9e5821f38
—Aggregate adapter SHA-256: dde649ca3166881b675a3db093ee273c6186e3d8e801c8491fac0a2da03e58f7
—Adapter weights SHA-256: 885fc29fcb4cf55c280dc085fdb0a40f40d6b946fee400dd5e4ed3459fe6334f

Only the repository-owned ReceiptAgent curriculum was admitted. The A11oy Brain corpus was excluded from gradients because row-level rights and provenance admission have not passed.

Held-out acceptance

The exact saved adapter was reloaded on the same GPU and evaluated against committed, digest-pinned held-out files:

Gate	Result
JSON contract-valid drafts	5 / 5
Adversarial refusals	6 / 6

These are raw MEASURED acceptance counts for a small preregistered gate. They are not a broad benchmark, do not establish factual accuracy, and do not make the adapter autonomy-eligible.

Evaluation report SHA-256: 0852fe55716da7b5fddf2340a00dd632c34d551096c685bd84923eb164f2a420.

Evidence boundary

The public, proposal-only release is available at `SZLHOLDINGS/szl-receiptagent-qwen35-0.8b-v2`. Its immutable evidence revision is 7a0d9efdebe92ea3b5b26b97cc722c6b5afde621.

The release includes:

—an owner-signed training receipt;
—an owner-signed evaluation receipt chained to the training receipt;
—the public Ed25519 verification key;
—source, dataset, adapter, and report digests; and
—an owner-signed post-publication receipt after the Hub bytes were independently fetched and hashed.

All 26 intended files (63,499,318 bytes) were fetched from the immutable Hub revision and matched their source or generated-artifact digest. The Hub-managed .gitattributes file is accounted for separately. The adapter weights matched SHA-256 885fc29fcb4cf55c280dc085fdb0a40f40d6b946fee400dd5e4ed3459fe6334f.

The owner-signed evidence chain binds source commit 2a5f9cd98503923a49c34e4ef3d3f92e40ca1386 and the durable qualification source-bundle SHA-256 b7b0eb703062981636d7cca41ed9ecaa8ca87a0ba51bd515cca721a6df0f5f8a to:

—training receipt canonical SHA-256: 8c5e74892b8a9933ffea074f92b7081f360634a1e8397a249f07a3efd06ca433
—evaluation receipt canonical SHA-256: f3b79e4b3aed7359c56ad33b72c7c4f6a4a517166d097a220fa8f3d3fa5402ff
—Ed25519 key ID: e7f01810aaa97394

The publication receipt canonical SHA-256 is 51444e7a8d6a6556ec848641620fbbea6b47e8363254aaf50d4ccdd38d6210cc. The evidence chain is valid and the bounded publication gate has passed. Autonomy eligibility remains false.

Immutable Hub runtime readback

The adapter was loaded from the exact Hub revision above and executed on an NVIDIA GeForce RTX 5050 Laptop GPU:

—28 generated tokens in 16.274 seconds;
—peak reserved GPU memory: 1,004,535,808 bytes;
—generated-output SHA-256: 909cfdd105fc78b9d665ba5dffd12e0a0feff554ba9d1c42415aeca29e200396;
—required refusal prefix observed: yes.

This is one immutable-revision GPU inference receipt. It is not a broad quality, safety, factual-accuracy, autonomy, or third-party benchmark claim.

Limitations

—Narrow synthetic curriculum.
—Small held-out set.
—Proposal-only behavior.
—No independent third-party evaluation.
—No ground-truth retrieval or autonomous execution.
—English-dominant evaluation.
—The validating controller remains mandatory.