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/SZL-Reasoning-0.8B-v1")model = AutoModelForCausalLM.from_pretrained("SZLHOLDINGS/SZL-Reasoning-0.8B-v1", 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.
Model overview
Table | |
|---|
| Developer | SZL Holdings |
| Architecture | qwen3_5 (config.json; text 24 layers, hidden 1024, plus vision tower) |
| Parameters | 873,438,784 (Hub safetensors metadata) |
| Base model | not yet recorded by a receipt (architecture matches Qwen/Qwen3.5-0.8B, apache-2.0) |
| Artifact | full weights, 1 safetensors shard (1,746,882,752 bytes) |
| Release date | 2026-11-06 (Hub createdAt) |
| Card generated | 2026-11-19 from repo revision 5ec45d7976f836f1fd04a02b6a742f61e0bb3d40 |
| Source | not mapped to a GitHub source in the estate registry |
Intended use
Research on estate-specific reasoning and confabulation behaviour at 0.8B. Not for answering questions about SZL systems without constrained decoding and an external evidence check.
Out of scope
Any production, factual or customer-facing use.
Quick start
from transformers import AutoModelForCausalLM, AutoTokenizerm = AutoModelForCausalLM.from_pretrained("SZLHOLDINGS/SZL-Reasoning-0.8B-v1", revision="5ec45d7976f836f1fd04a02b6a742f61e0bb3d40")
Evaluation
Table with columns: Battery, Result, Receipt, Date| Battery | Result | Receipt | Date |
|---|
| Fleet confabulation sweep (6 probes, greedy) | abstained 1, answered 5 · verdict FAIL | receipts/eval/fleet-confabulation-sweep-20261108.json sha256 9a16903f…a8ebb3 | 2026-11-08 |
Estate battery v2 (3 answerable + 3 unanswerable, RTX 4090, Salad run szl-estate-battery-v2) | 0 / 6 abstained or refused | frontier-54-runs receipts/estate-battery-szl_reasoning_0_8b_v1.json sha256 b3c22622…fd0d | 2026-11-15 |
Standard benchmarks (GSM8K, MATH500, GPQA, IFEval): not yet measured.
Training
Table | |
|---|
| Recipe | not yet recorded by a receipt |
| Data | not yet recorded by a receipt |
| Steps / epochs | not yet measured |
| Final train loss | not yet measured |
| Hardware | not yet recorded |
Receipts
Table with columns: File, sha256, Signature| File | sha256 | Signature |
|---|
| receipts/eval/fleet-confabulation-sweep-20261108.json | 9a16903f6943917a690141d8b2a2fa0bdc485bd8981a44518e53ed405fa8ebb3 | unsigned |
| frontier-54-runs: receipts/estate-battery-szl_reasoning_0_8b_v1.json | b3c22622cfd5756f3afd4907fd77742a3079e182bc1a583a0af75ba4efc6fd0d | unsigned |
Limitations
- Fails the estate confabulation battery: it answers questions about SZL systems it cannot know, under free decoding (receipt honest_note: the floor at 0.8B–4B is a decoding property; grounding must be enforced at inference).
- No training provenance is verifiable from this repository.
- No general capability, benchmark, safety or production-readiness claim is made.
Licence
Card declares apache-2.0. Upstream licence confirmation is pending a training receipt naming the base.