Model overview
Table | |
|---|
| Developer | SZL Holdings |
| Architecture | Qwen3.5 (qwen3_5, config.json) |
| Parameters | 873,438,784 (safetensors metadata) |
| Base lineage | Qwen/Qwen3.5-0.8B (apache-2.0); training runtime unsloth/Qwen3.5-0.8B |
| Artifacts | merged weights (root) + LoRA adapter (adapter-unsloth/) |
| Card generated | 2026-11-19 from repo revision 89c5c22c56d42790313b696cfd2e4b3648c4e30a |
Intended use
Research on small governed drafters that emit proposals in a fixed JSON contract and refuse out-of-scope requests. Outputs are proposals for a controller to validate.
Out of scope
Any production or autonomous use. The measured gate failed, so this checkpoint should not be relied on for valid JSON drafts or for refusals.
Quick start
from transformers import AutoTokenizer, AutoModelForCausalLMtok = AutoTokenizer.from_pretrained("SZLHOLDINGS/chaski")model = AutoModelForCausalLM.from_pretrained("SZLHOLDINGS/chaski", 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 |
|---|
| JSON draft contract | 0 / 5 | eval_measured.json | 2026-08-28 |
| Adversarial refusal | 2 / 6 | eval_measured.json | 2026-08-28 |
Method (receipt field): in-process greedy generation on the merged shard, szl-forge PR 63 gates; parent Hub sha 1c55df86…. Standard benchmarks (MMLU-Pro, GPQA, IFEval, BFCL), energy and process-peak memory: not yet measured.
Training
Table | |
|---|
| Dataset | SZLHOLDINGS/szl-1-doctrine-sft file szl_dataset.jsonl, sha256 ddc5594b…a243, 45 rows |
| Recipe | LoRA r16 / alpha 32, lr 2e-4 constant with warmup (6), adamw_8bit, response-only loss, seed 11, 64 max steps |
| Result | train loss 1.7839 at epoch 2.8, runtime 161.6 s (training loss, not an evaluation) |
| Hardware | not yet measured |
Receipts
All hashes were computed on 2026-11-19 from the files at the revision above. szl-source-binding.json declares this README is projected from szl-holdings/szl-forge; this card was written directly on the Hub on 2026-11-19 and should be ported back to that source.
Limitations
- The measured gate failed: 0/5 valid JSON drafts and 2/6 adversarial refusals.
- 45 training rows is a very small curriculum.
- No held-out generalisation benchmark exists.
- 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_chaski_2026, title={chaski}, author={SZL Holdings}, year={2026}, howpublished={\url{https://huggingface.co/SZLHOLDINGS/chaski}}}
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.