Model overview
Table | |
|---|
| Developer | SZL Holdings |
| Architecture | Qwen3.5 (qwen3_5, config.json) |
| Parameters | 852,985,920 (safetensors metadata) |
| Base lineage | SZLHOLDINGS/szl-receiptagent-qwen35-0.8b-v2 (LoRA) on Qwen/Qwen3.5-0.8B (apache-2.0) |
| Merge receipt | not in this repo; the adapter repo's salvaged-merge/merge_receipt.json records a merge |
| Card generated | 2026-11-19 from repo revision 19f86aec2a65eb549d464615cffce9cb049d5c33 |
Intended use
Research comparison of merged vs adapter behaviour.
Out of scope
Any use that depends on the model saying "I don't know". On the estate battery it produced confident answers, including fabricated specifics, to all 6 questions it could not know.
Quick start
from transformers import AutoTokenizer, AutoModelForCausalLMtok = AutoTokenizer.from_pretrained("SZLHOLDINGS/szl-receiptagent-qwen35-0.8b-v2-merged")model = AutoModelForCausalLM.from_pretrained("SZLHOLDINGS/szl-receiptagent-qwen35-0.8b-v2-merged", 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))
Run check (2026-11-19, transformers 5.22.2, torch 2.14.1 (CPU)): the repository loads and generates text from a plain-text question. This is a smoke test, not an evaluation. The tokenizer ships a chat template. In transformers 5.x apply_chat_template returns a BatchEncoding, so pass return_dict=True and **inputs as shown (the earlier ids.shape[1] form fails on 5.x).
Evaluation

Chart data and sources: assets/chart-data.json. Every bar is a count from the receipt named in the table below.
Table with columns: Battery, Result, Receipt, Date| Battery | Result | Receipt | Date |
|---|
| Estate battery v1 (abstain on unanswerable) | 0 / 6 abstained or refused | receipts/eval_battery.json | — |
The receipt-agent contract eval of the adapter (5/5, 6/6) was run on the adapter, not on these merged weights; merged-weight contract eval: not yet measured. Standard benchmarks (MMLU-Pro, GPQA, IFEval, BFCL), energy and process-peak memory: not yet measured.
Provenance
Table | |
|---|
| Training | none in this repo; merge of the v2 adapter (see the adapter card) |
Receipts
Table with columns: File, sha256, Signature| File | sha256 | Signature |
|---|
| receipts/eval_battery.json | b59ce231899086f45cd6fdc3803c5ec065dc52eb47607c455cf5277dbf49fdbf | unsigned |
All hashes were computed on 2026-11-19 from the files at the revision above. No source binding file in this repo.
Limitations
- 0/6 abstentions on the estate battery: it fabricates answers to unknowable questions.
- Merge provenance is recorded in the adapter repo, not here.
- No contract eval on the merged weights.
- Safety evaluation (toxicity, jailbreak and red-team suites): not yet evaluated.
Licence
Apache-2.0, inherited from the base model (apache-2.0). Keep the upstream LICENSE and notice.
Citation
@misc{szl_szl_receiptagent_qwen35_0_8b_v2_merged_2026, title={szl-receiptagent-qwen35-0.8b-v2-merged}, author={SZL Holdings}, year={2026}, howpublished={\url{https://huggingface.co/SZLHOLDINGS/szl-receiptagent-qwen35-0.8b-v2-merged}}}
SZL mark from szl-holdings/szl-brand kit/logos/orbit-v2/szl-orbit.svg (CC BY 4.0). Khipu cords in the banner encode receipt counts (knots = decimal digits). The Λ is identity artwork; Conjecture 1 remains open.