Training and artifact
- Fresh base
Qwen/Qwen3.5-9B, revision c202236235762e1c871ad0ccb60c8ee5ba337b9a.
- 3,900 complete federal Acts, expanded into 10,960 whole/owned-unit chats:
3,231 whole-chat fits and 669 complete reference-guided split plans.
- Two epochs, 21,920 chat presentations, 5,480 optimizer updates;
microbatch one, accumulation four, learning rate
5e-6, seed 102301.
- All language parameters trained; unused vision frozen. FP32 masters/buffers,
BF16 autocast, factored FP32 Adafactor; no adapters or quantization.
- Reviewed-reference weight 1.0, silver 0.5, divided by owned-unit count and
globally normalized. Completion-only loss. Training context: 32,768 tokens.
- Uploaded artifact is the unchanged native FP32 checkpoint, not a BF16
re-export. BF16 loading is an inference-time cast. No optimizer state is uploaded.
- Native execution corpus SHA-256:
9074e0f8ce109e88883f5c6323d323020ae7852ec053134ecd3bd1162ab7ecd7.
Training targets are source-reference projections, not expert semantic gold.
Reference-guided training subdivision is not a qualified source-only inference
splitter or a whole-Act assembly method. See native manifests and provenance.
Evaluation comparison: Acts model vs Sol
Source-size-qualified cohort: 100 CT + 100 federal training-unseen Acts, each
complete chat-templated input below 34,000 Qwen tokens (maximum 33,947).
Selection uses source hashes and input size, not successful outputs. Training
identity and exact/whitespace-normalized source separation were checked. CT Acts
were previously inferred; this is not a fresh blind CT test. No fragments,
truncation, JSON repair or failed-case replacement. All failures remain in the
original 200-document denominator.
Both models received the same frozen whole-Act messages. This table measures
strict output-contract adherence, not semantic accuracy.
Table with columns: Measure, XLL Acts Qwen3.5-9B, GPT-6.1 Sol| Measure | XLL Acts Qwen3.5-9B | GPT-6.1 Sol |
|---|
| CT: nontrivial span-valid trees | 55/100 (55%) | 0/100 |
| Federal: nontrivial span-valid trees | 88/100 (88%) | 0/100 |
| All: nontrivial span-valid trees | 143/200 (71.5%) | 0/200 |
| Other invalid outputs (JSON/tree) | 55 | 200 |
| Output-cap failures | 2 | 0 |
| Recorded outcomes / original denominator | 200/200 | 200/200 |
Comparison limitation: the frozen trained prompt does not enumerate allowed
tags. Sol's zero is a strict-contract result, not zero legal understanding;
this is not an optimized Sol prompt or an equal-output-budget comparison.
Inspect the paired results,
Acts summary,
Sol summary, and
Sol validation audit.
Original predictions are preserved for
Acts and
Sol; the
frozen identical requests identify the cohort.
RTX 5090, BF16, batch one, greedy, thinking disabled, flash SDP; 65,536-token
context with 16,384 output tokens, one-hour per-case deadline. This exceeds the
trained 32K context: longer-context quality is unqualified. There were no campaign
runtime stops; 200 outcomes were recorded. These are geometry/format checks,
not exact-tree or semantic accuracy. Independent qualified references are absent.
Sol (openai/gpt-6.1-sol) received identical trained-prompt messages for the same
200 whole Acts without an added output cap: 169 parsed outputs failed on unknown
tags and 31 were invalid JSON; zero passed the strict contract. The trained
prompt does not enumerate the allowed tags, and output budgets differ. This is
not an optimized Sol comparison or evidence that Sol has zero legal understanding.
No Sol outputs are used as gold. Frozen cohort, outputs, receipts and comparison
caveats are included under evaluation/.
Required interface
Use losslessly indexed Acts text, not raw text or XML. Offsets refer to the
original unnormalized Unicode text, end exclusive; never to the indexed message.
Use PROMPT_TRAINED.txt byte-for-byte for whole Acts (SHA-256
57766ac499dd8ab8b1d20901fd14447af2c2477d57433c2d4c11f2486618ea3b).
Allowed structural tags are recorded in OUTPUT_CONTRACT.json; this vocabulary
was not appended to the frozen benchmark messages. Fragment training uses a
different system prompt preserved in the exact execution dataset; it does not
qualify fragment assembly.
import json, refrom pathlib import Pathimport torchfrom huggingface_hub import HfApi, hf_hub_downloadfrom transformers import AutoTokenizer, AutoModelForImageTextToText repo = "thurstonpowers/xll-acts-qwen3.5-9b"revision = HfApi().model_info(repo, revision="acts-20261118").shasystem = Path(hf_hub_download(repo, "PROMPT_TRAINED.txt", revision=revision)).read_text()raw = Path("act.txt").read_bytes().decode("utf-8")parts = [f'{m.start()}:{m.end()}|{json.dumps(m.group(), ensure_ascii=True)}' for m in re.finditer(r'\s+|\w+|[^\w\s]', raw) if m.group().strip(' ')]indexed = f"Document type: acts. Text length: {len(raw)} Unicode characters.\n" + "\n".join(parts)tokenizer = AutoTokenizer.from_pretrained(repo, revision=revision, trust_remote_code=False)ids = tokenizer.apply_chat_template([ {"role": "system", "content": system}, {"role": "user", "content": indexed}], tokenize=True, add_generation_prompt=True, enable_thinking=False, return_dict=False)max_new_tokens = 16384if len(ids) + max_new_tokens > 32768: raise ValueError("Whole input plus output reserve exceeds trained context; do not truncate")model = AutoModelForImageTextToText.from_pretrained(repo, revision=revision, dtype=torch.bfloat16, attn_implementation="sdpa", trust_remote_code=False).to("cuda").eval()input_ids = torch.tensor([ids], device="cuda")with torch.inference_mode(): result = model.generate(input_ids=input_ids, attention_mask=torch.ones_like(input_ids), max_new_tokens=max_new_tokens, do_sample=False, use_cache=True, pad_token_id=tokenizer.pad_token_id or tokenizer.eos_token_id)text = tokenizer.decode(result[0, len(ids):], skip_special_tokens=True)prediction = json.loads(text) # Parsing is not span validation or semantic review.
Evidence runtime: Transformers 5.15.1, Torch 2.13.0+cu130 on RTX 5090; native
training used the pinned NVIDIA runtime recorded in provenance/. Compatibility
with arbitrary library versions or CPU runtimes is not established.
License and provenance
Apache-2.0 base-model license is included as LICENSE; see NOTICE. This license
does not relicense source texts or certify dataset rights. SHA256SUMS.json
binds weights, tokenizer, prompts, evaluation and native completion/training
receipts. Local paths in historical receipts are provenance, not download
dependencies. No website deployment or new training follows from this release.