SZLHOLDINGS/SZL-Forge-1.5B-ReceiptAgent

TEXT GENERATIONConcurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 13, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

SZL-Forge-1.5B-ReceiptAgent by SZL Holdings is a 1.5 billion parameter, governed, proposal-only agent fine-tuned from Qwen/Qwen2.5-1.5B-Instruct. This model specializes in emitting evidence-bound, approval-gated decision drafts as JSON, ensuring it never finalizes or executes actions. It is designed to propose drafts conforming to a specific ReceiptAgent output schema, refusing to overstep its boundary as a proposer within a controller system.

Loading preview...

SZL-Forge-1.5B-ReceiptAgent Overview

SZL-Forge-1.5B-ReceiptAgent is a specialized 1.5 billion parameter model developed by SZL Holdings, fine-tuned from Qwen/Qwen2.5-1.5B-Instruct. Its core function is to act as a governed, proposal-only agent, generating evidence-bound decision drafts in JSON format. A key differentiator is its strict adherence to a controller boundary: it proposes drafts but never finalizes, executes, or fabricates information, and will refuse if prompted to overstep these limits.

Key Capabilities

  • JSON Draft Generation: Emits single JSON drafts conforming to the ReceiptAgent output schema, with decision=DRAFT, approvalRequired=true, executed=false, and provenance=MODEL_PROPOSED.
  • Evidence Binding: Includes at least one cited evidence source with an honest label (e.g., MEASURED, REPORTED).
  • Refusal Mechanism: Designed to refuse prompts that ask it to act autonomously or finalize decisions, maintaining its role as a proposer.
  • Verifiable Provenance: Every capability claim is backed by ed25519 owner-signed receipts (training and evaluation) that are hash-chained and independently re-verified by the Alloy backbone.

Training and Evaluation

The model was trained using QLoRA SFT with response-only loss masking and refusal oversampling, based on a deterministic, schema-validated synthetic curriculum. Evaluation on a held-out curriculum shows 100% draft-conformance (5/5 schema-valid drafts) and 100% adversarial-refusal (6/6 correctly refused oversteps).

Intended Use

This model is ideal for proposing governed, evidence-cited decision drafts within a human-in-the-loop controller system (e.g., Alloy). It is explicitly not intended for autonomous execution, finalizing actions, or as a source of ground-truth numbers.

Finetunes
1 models