Charts and quick start

Chart data and sources: assets/chart-data.json. Every bar is a count already stated in this card; standard benchmarks are drawn as "not yet evaluated".

Sizes are Hub file sizes on 2026-11-19. They say nothing about quality.
Quick start
from peft import PeftModelfrom transformers import AutoTokenizer, AutoModelForCausalLMbase = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-0.8B", torch_dtype="auto")model = PeftModel.from_pretrained(base, "SZLHOLDINGS/triage-0.8b-v2-trained")tok = AutoTokenizer.from_pretrained("SZLHOLDINGS/triage-0.8b-v2-trained")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), peft 0.22.0): the code path of this snippet (adapter attach, 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.
- The charted counts come from batteries of 3 probes: small samples, a gate check and not a benchmark.
- This repository holds adapter weights; they need the base model
Qwen/Qwen3.5-0.8B and are not a standalone model.
Quick facts
Table | |
|---|
| Base model | Qwen/Qwen3.5-0.8B |
| Type | LoRA adapter (see adapter_config.json for rank/alpha) |
| License | apache-2.0 |
| Downloads | 23 |
| Evidence dataset | SZLHOLDINGS/frontier-54-runs |
Measured evaluation battery (receipt-bound)
Table | |
|---|
| Receipt kind | szl.estate-battery/mass-training-batch2 |
| Measured on | NVIDIA RTX A6000 — node 0056-kci3-sma6-prxmx200114 |
| Receipt timestamp | 2026-11-14T01:40:16.657931+00:00 |
| Score | 3/6 |
| Energy | see receipt |
| Chain receipt | benchmark/triage-0.8b-v2-trained_battery.json |
| Mirrored | receipts/eval_battery.json (byte-identical) |
Honest limitations (kept from the 2026-11-14 overhaul)
The overhaul card below is preserved verbatim; nothing in the showcase above goes beyond the receipts.
Measured status: EVALUATION BATTERY BOUND (2026-11-14)
- An estate evaluation battery receipt (kind
szl.estate-battery/mass-training-batch2) naming this
repository was measured on NVIDIA RTX A6000 (node 0056-kci3-sma6-prxmx200114) at 2026-11-14T01:40:16.657931+00:00 and published in the estate
evidence dataset SZLHOLDINGS/frontier-54-runs at benchmark/mass-battery-triage_0_8b_v2_trained.json (dataset commit febc1eb24f8f8c0ec90c688d9b81a007597f24c1).
It is mirrored byte-identically in this repository at receipts/eval_battery.json
(sha256 c45ce6f8550f81a80be4ac801d22b619f3a288554ee52c1008204216fd55665b).
- Battery result (state
MEASURED): 3/6 — 3/3 answerable probes passed; 0/3 unanswerable
probes were correctly abstained or refused.
- Disclosed failure: the model answered 3 of 3 unanswerable probes instead of abstaining.
This is a measured abstention failure, not a capability claim.
- Binding scope: the receipt names this repository and its training-lane run window; it does not
embed this adapter's byte-level digest, so the binding is by repository identity, not a
weight hash.
- No training receipt (step count, loss curve, hardware, duration) was published with this
artifact. Training provenance remains not independently verifiable from this repository.
- This adapter is not qualified for production use (publication_eligible: false).
Curated repository facts (curated-card batch-4 2026-11-16)
- Observed Hub revision immediately before this edit:
aab4ec2aadf33e7911c4dc1659b5bc96b3afbd2a. The prior card remains in revision history.
- Observed repository creation date:
UNAVAILABLE.
- Observed repository contents: 9 files (1 weight files, 2 config/tokenizer files, 2 receipt files, 3 other files).
Receipts present:
receipts/eval_battery.json, receipts/receipt_visual.svg.
- License tag observed verbatim in this card's frontmatter:
apache-2.0. This edit changed no license and asserts none beyond the observed tag.
- Source mapping (estate asset registry): GitHub szl-holdings/szl-forge observed at HEAD
6c21496eaf346f8b2184b711335987c007a69ab0 (last push 2026-11-16T00:47:09Z); registry method card.
- Atelier classification: none recorded in the estate manifest. No class-based claim is made on this card.
- Eval battery receipt bound:
receipts/eval_battery.json (sha256 c45ce6f8550f81a80be4ac801d22b619f3a288554ee52c1008204216fd55665b), kind szl.estate-battery/mass-training-batch2, score 3/6 (MEASURED), measured 2026-11-14T01:40:16.657931+00:00 on NVIDIA RTX A6000 (node 0056-kci3-sma6-prxmx200114).
Battery verdicts: 3/3 answerable probes passed; 0/3 unanswerable probes correctly abstained or refused.
Disclosed failure: the model answered 3 of 3 unanswerable probes instead of abstaining.
- Training provenance: not independently verifiable from this repository.
- Hub reads at manifest generation (2026-11-14): None.
- This edit adds observed repository facts only. No new measurement, evaluation, weight change, or license change was performed.
Card history
- On 2026-11-14 an off-route automated job overwrote this repository card with a marketing
template containing fabricated per-model claims. That text was not evidence and has been
replaced by this honest minimal card on 2026-11-14.
- Estate record: szl-holdings/szl-forge docs/model-card-overhaul-20261114
(schema, manifest, and rules for this overhaul)
License
No license is declared in this repository metadata. The adapter is derived from Qwen/Qwen3.5-0.8B; the base model repository and its license govern redistribution and use of this adapter. No license is asserted here because none is evidenced (estate rule: licenses are never invented).
This card is part of the 2026-11-14 estate model-card overhaul and follows its rules: no metric
without a verified receipt binding; unverified cards carry an explicit not-measured gap;
licenses are never invented.
License (derived 2026-11-14)
license: apache-2.0 is derived, not invented: this repository is a LoRA adapter over
Qwen/Qwen3.5-0.8B (recorded in adapter_config.json: base_model_name_or_path), whose model card declares
license: apache-2.0 at revision 2fc06364715b967f1860aea9cf38778875588b17 (read via the HF API on 2026-11-14).
- This repository contains adapter weights produced by the estate's own training runs;
no upstream base-model weights are redistributed here.
Evaluation battery addendum (receipt-bound 2026-11-14)
Bound in the full-estate receipt-binding pass (owner direction 2026-11-14: solo build; bind every published eval receipt, all models public and private).
The battery receipt backing this card honesty score is published on the receipts chain and is mirrored byte-identical in this repository.
- Receipt (chain):
benchmark/mass-battery-triage_0_8b_v2_trained.json in SZLHOLDINGS/frontier-54-runs (dataset commit 13247cfb85d5)
- Mirrored in-repo:
receipts/eval_battery.json (sha256 c45ce6f8550f81a80be4ac801d22b619f3a288554ee52c1008204216fd55665b)
- Receipt self-digest (
receipt_sha256): 8e3bcf5ae782ae92baf05e601aedb7fea8bb30223ea7528ba1d07d0b214c1188)
- Measured: score 3/6, state MEASURED, ts 2026-11-14T01:40:16.657931+00:00, GPU NVIDIA RTX A6000, node 0056-kci3-sma6-prxmx200114, kind
szl.estate-battery/mass-training-batch2, base Qwen/Qwen3.5-0.8B
- Per-probe verdicts: a1: PASS; a2: PASS; a3: PASS; u1: FAIL (fabricated); u2: FAIL (fabricated); u3: FAIL (fabricated)
- Binding scope: identity binding to the receipt that names this repository; it carries no adapter byte digest.
- Badge note: where a receipt badge SHA in this card differs from the receipt self-digest above, the chain receipt is authoritative.
- Not covered: training-configuration specifics (steps, learning rate, rank/alpha) are not covered by any published training receipt for this model; treat as unverified until a training receipt is bound.
Citation & reuse
Cite the estate evidence dataset and this repository. Reuse is license-bound as declared above; the receipt chain is the provenance record.