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/brain-navigator-r2")model = AutoModelForCausalLM.from_pretrained("SZLHOLDINGS/brain-navigator-r2", 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.
What it does
LoRA (bf16, r=16, α=32) proposing NAVIGATE/ABSTAIN over supplied public handles. 0 raw private-graph nodes admitted to gradients (per training receipt).
What it does not do
- Encode or grant private 9,464-node graph access
- Transfer named-N generate scores to repaired merged checkpoint automatically
- Overwrite
SZL-Khipu-1.5B / GGUF BrainNavigator lineage
- Authorize autonomy
Intended use
- Research PEFT attach at pinned revision
- Comparing named-N generate vs later held-out gate FAIL
- Controller-bound retrieval-plan experiments on public projections
Out of scope
- Private-graph content resolution
- Production promotion from Hub file presence
- Combining incompatible suites into one score
- Laptop-lora sibling as canonical
Training data
No Hub datasets: id historically asserted as sole train mix; this wave links adjacency only.
Evaluation (MEASURED or UNKNOWN)
Table with columns: Suite, Result, Class| Suite | Result | Class |
|---|
| Named-N generate (eval_report.json) | retrieval 5/5 · abstention 6/6 · parse fail 0 · hallucinated citations 0 | MEASURED (owner-run; not public leaderboard) |
| Post-publication gate (2026-09-15) | 2/5 held-out · overall FAIL · refusalPassed=false | MEASURED |
publication_eligible / autonomy_eligible | false | MEASURED policy |
| Merged checkpoint behavioral eval | — | UNKNOWN (merge ≠ new eval) |
Energy / BOM
Table with columns: Field, Status| Field | Status |
|---|
| Energy | UNAVAILABLE |
| BOM | UNAVAILABLE on card body (add bom/model-bom.cdx.json in a later packaging wave if/when present) |
| DSSE | Prefer verify linked gate receipts; unsigned local notes stay honest UNSIGNED |
Usage (research)
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizerbase_id = "Qwen/Qwen3.5-0.8B"base_rev = "<BASE_PIN>"adapter = "SZLHOLDINGS/brain-navigator-r2"adapter_rev = "ad510c4b05429acb862b216718f098ace6d05af3"tok = AutoTokenizer.from_pretrained(base_id, revision=base_rev)base = AutoModelForCausalLM.from_pretrained(base_id, revision=base_rev, device_map="auto")model = PeftModel.from_pretrained(base, adapter, revision=adapter_rev)
Do not overwrite canonical serve tags (chaski, SZL-Khipu-1.5B, szl1).
Source of truth
- Hub evidence + gate receipt
- Historical authoring card under szl-forge
brain-navigator-r2/card/
Limitations
- Later gate FAIL supersedes any casual “green” reading of named-N
- Train loss /
evals=none-this-run on train step ≠ held-out quality
- Candidate / laptop-lora ids are clutter
Doctrine pin
- Doctrine: v11 LOCKED
- Λ (Lambda): Conjecture 1 — advisory, never a theorem
- Evidence law: MEASURED or null / UNKNOWN / UNAVAILABLE — never fabricate energy, eval, or DSSE
- Motto: Control before action. Evidence after.
License
Apache-2.0. See LICENSE in-repo when present.
Upgrade wave note
NVIDIA-style card upgrade draft · 2026-11-04T17:41Z · clean-room silhouette (DSSE receipts + MEASURED-or-null energy).
Does not authorize train, serve, or autonomy. Does not overwrite sibling SKUs.
The estate
Provenance note: duplicate weights
The listed weight files in this repository are byte-identical (same LFS SHA-256; checkpoint-* excluded) to files in another repository: SZLHOLDINGS/brain-navigator-r2-laptop-lora@64531aaa4399 (shared: adapter_model.safetensors) (checked 2026-11-12 at e0dfb5b632dc). Identical bytes are one artifact published under more than one name; this repository is not independent evidence of a separate training run, and any results stated for it apply to that same artifact. Existing result lines above are left unchanged.
Measured evidence: TRAINING RECEIPT BOUND (2026-11-14)
- Receipt file:
training_receipt.json (sha256 48d2bd16582c…)
- Receipt facts: {"kind": "szl-brain-navigator-r2-training-receipt", "schema": "szl.frontier-training-run/v1", "v": 1, "artifact": "SZLHOLDINGS/brain-navigator-r2", "sku": "BRAIN-NAVIGATOR-R2", "canonical_base": "Qwen/Qwen3.5-0.8B", "base_model": "Qwen/Qwen3.5-0.8B", "qlora": false, "load_in_4bit": false, "load_in_16bit": true, "quant": "bf16-lora", "lora_r": 16, "lora_alpha": 32, "seed": 11, "num_train_epochs": 3, "warmup_steps": 6, "learning_rate": 0.0002, "lr_scheduler_type": "constant_with_warmup", "optim": "adamw_8bit", "response_only_loss": true, "max_seq_length": 2048, "dataset_file": "train/train.jsonl
- Note: receipt file already present in-repo at audit; card binding added 2026-11-14. Re-public earned per estate doctrine (receipts earn visibility).
Curated repository facts (curated-card batch-3 2026-11-16)
- Observed Hub revision immediately before this edit:
3dde90f7f8ae4d4f594518f92114f36cee0b5a19. The prior card remains in revision history.
- Observed repository creation date:
UNAVAILABLE.
- Observed repository contents: 18 files (2 weight files, 4 config/tokenizer files, 11 other files).
- 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: none in this repository. No evaluation result, benchmark, or model-index metric is claimed in this edit.
- Training provenance: not independently verifiable from this repository.
- Hub reads at manifest generation (2026-11-14): 4.
- This edit adds observed repository facts only. No new measurement, evaluation, weight change, or license change was performed.