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/chaski-5050")model = AutoModelForCausalLM.from_pretrained("SZLHOLDINGS/chaski-5050", 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.
Before you use it
- The later held-out gate recorded 1/5 and overall FAIL. Publication and autonomy eligibility remain false.
- Lab loading remains forbidden; root adapter metadata contains a nonportable owner-local base path and runtime loading is unverified.
- Do not substitute Chaski or R2 evidence. Adapter results require explicit loader-class and tensor-coverage evidence.
Evidence passport
Dated input observation: 2026-11-13T22:16:59.048614+00:00. Existing research limits remain in force.
Inspect the immutable repository snapshot.
Table with columns: Observation, Scope| Observation | Scope |
|---|
| Repository revision | 6bd12f97b3d19415c018473f9ff0427ad9343b1a |
| Historical README retrieval | HTTP 200; not a readback of this proposed panel |
| Recognized artifact filenames | 3 filename hints; not tensor verification |
| Structured records retrieved | 3; bounded selection, not an exhaustive evidence census |
Historical input README SHA-256: 0d7f289daec1e217a41c9ca55c651b0fe51eb302aaafeafcccc09124d0923987.
Reviewed authoring baseline. Existing publisher: publish-chaski-card.yml. Source changes still require normal review and exact deployed-byte readback.
Inspect the records
Interpretation
Documentation and retrieved-record integrity only. No new training, model evaluation, weight-byte verification, independent signer identity, runtime qualification or promotion. Missing evidence is not a zero score.
Training loss is not held-out performance. A copied receipt does not qualify a quantization or merged checkpoint. Existing negative results, license and technical disclosures remain authoritative.
Artifact identity (noted 2026-09-30). Besides the LoRA adapter, this repository's root carries a merged full-precision checkpoint (model.safetensors + config.json, loadable with transformers) produced by the receipted CPU merge of this adapter into its declared base (merge_receipt.json). The adapter remains the artifact of record for every figure on this card; the merged bytes carry no separate held-out receipt and add no claim. Metadata-only note.
Loader-class boundary (noted 2026-09-30). The adapter tensors are keyed for the multimodal module layout (base_model.model.model.language_model.layers.*, the class Qwen3_5ForConditionalGeneration / AutoModelForImageTextToText). Under transformers 5.18, AutoModelForCausalLM instantiates Qwen3_5ForCausalLM (model.layers.*); PEFT then applies 0 of the 192 adapter tensors and emits only a warning, so the result is the bare base model (observed and receipted on the sibling SZLHOLDINGS/chaski-r2 adapter, which shares this exact key layout: byte-identical base outputs on the held-out prompts). Any run that reports a score for this adapter must show adapter key coverage (192/192) and the loader class it used; a record without those fields does not establish that the adapter was applied. Evidence: szl-holdings/szl-forge tools/geh_v8/evidence_sandbox/chaski_probe/ (CPU replay receipts, #444). Metadata-only note; it authorizes no load, changes no artifact, and adds no claim.
Chaski-5050
QUARANTINE. Research residue. Root adapter metadata contains a
nonportable owner-local base path. Runtime loading has not been verified.
This bf16 LoRA experiment is separate from SZLHOLDINGS/chaski and
Chaski-R2. Publication and autonomy eligibility remain false.
Lab load is forbidden. The card does not authorize artifact changes or
model promotion.
Card and metadata review snapshot: 8c9d4782b76e55377e00fba249c6938000dca423.
Small source and receipt files were read; no model weights were downloaded,
no inference or tensor comparison was performed, and no current artifact
was qualified.
Table with columns: Form, Observed metadata, Qualification limit| Form | Observed metadata | Qualification limit |
|---|
| Root LoRA adapter | adapter_model.safetensors and adapter_config.json; historical card reports 25,587,104 bytes and raw-file SHA-256 620b3488fac2ebc6518090424de5b3c6a182293cf52dfd5bd9f886f54aef0df5 | The root config embeds an owner-local absolute base path and names Qwen3_5ForConditionalGeneration in auto_mapping. Loading is unverified. |
adapter-unsloth/ | Additional adapter and processor metadata are listed | Lineage and evaluation applicability were not independently established here; this is not interchangeable with the root adapter by assumption. |
| Root merged checkpoint | model.safetensors, config.json, and merge_receipt.json are listed | A separate runtime form; file presence or a merge receipt does not establish behavioral qualification. |
The 2026-09-14 merge repair record
reports replacing an earlier base-copy merge, touching 96/96 modules with
maximum weight delta 0.0012716073542833328. This owner-reported repair
does not establish the quality, loading, or exact current identity of the
merged checkpoint.
Dated training and evaluation
The 2026-08-28 training receipt
records evals=none-this-run: no JSON/refusal gate ran in that training step.
Its train loss 2.228136855544466 and runtime 883.2224 seconds are
training measurements, not held-out evaluation scores.
A separate 2026-09-15 canonical gate receipt
at historical Hub revision 63ea80915ccb663b4aa589cbf2ce52a1dec574fe
reports 1/5 held-out and overall FAIL, with
heldout.refusal_no_regression=false. Its separate refusal command passed;
that does not change the overall failure. Production and publication
authorization are false. This receipt is a distinct owner-run suite;
its applicability to the current distributed forms was not independently
established in this review. No 5/5 or 6/6 qualification is claimed.
Historical recipe
The training receipt describes an owner-local RTX 5050 Laptop run, rather
than an HF Job: bf16 LoRA r=16, ฮฑ=16, seed 11, three epochs, batch size 1,
accumulation 4, maximum sequence length 2048, learning rate 2e-4, and
adamw_8bit. The dataset is SZLHOLDINGS/szl-1-doctrine-sft, 41 rows,
JSONL SHA-256 ddc5594bfb1c78449ba40a263f5ac41d21c896c3c7ed7346341c7c080611a243.
The recorded stack is Unsloth 2026.7.2, Transformers 5.5.0, and Torch
2.10.0+cu128. copied_live_chaski_weights=false is the receipt's declaration.
These records are preserved; this card review reran none of them.
Research loading boundary
The former generic AutoModelForCausalLM example was not validated against
the root adapter's multimodal wrapper metadata or the coexisting merged
weights. Treat this repository as research residue rather than a runnable
quickstart. Before loading, a source owner must select a particular form,
verify its compatible wrapper and immutable base revision, and resolve the
nonportable path under a separately reviewed artifact change. A card edit
does not sanitize the configuration or qualify a replacement loader.
A11OY-MINI has its own
artifact-specific legacy and R2 evidence. It inherits no training score,
gate result, or qualification from this card.
Apache-2.0 is declared in metadata. No standalone LICENSE is listed at
the reviewed revision; artifact license coverage was not independently
verified. A license addition requires source-owner review of that coverage.
Canonical card authoring source: chaski-5050/card/README.md.
The older chaski/README_5050.md reference is a separate recipe document,
not the card file selected by the canonical card publisher.
Doctrine v11 LOCKED 749/14/163. ฮ = Conjecture 1, advisory, never a theorem.
SZL Orbit artwork: source ยท CC BY 4.0. Software and artifact licensing remain declared above.
Publication restoration note (2026-11-14)
On 2026-11-14, an off-route automated job stamped the public SZLHOLDINGS Hub
model cards with a marketing template that replaced curated card bytes,
including per-model claims this artifact has not earned. This source card in
szl-holdings/szl-forge is the canonical reviewed card; that job did not alter
it. This note marks the card so the closed-profile publisher republishes the
reviewed bytes, banner, and source binding to the Hub with exact readback.
The incident and the per-model evidence manifest are documented in szl-forge
under docs/model-card-overhaul-20261114/. No qualification, evaluation,
weight, adapter, config, visibility, hardware, or runtime state changed by
this note.