Skip to content
Computer vision

XLL Acts Qwen3.5-9B

fine-tune for structural analysis of federal acts

2 min readupdated

An experimental checkpoint based on Qwen/Qwen3.5-9B (9.41B parameters), trained on US federal acts for text structuring tasks. The model takes indexed document text as input and formats its output according to a system prompt and output contract. It is not a production-ready build and does not guarantee legal or semantic accuracy, making it a fit for researchers parsing long official and legal documents.

Key features

  • Built on the recent Qwen3.5-9B model with a context window of up to 32,768 tokens.
  • Trained on 3,900 federal acts expanded into 10,960 chat presentations (2 epochs, 5,480 optimizer updates).
  • Trained using completion-only loss with a focus on structural markup.
  • Loads in native FP32 format with optional BF16 casting at inference, requiring no extra adapters or quantization.

What you can do

  • Run experiments on structuring and parsing lengthy text documents and official acts.
  • Build text-processing pipelines using the transformers library and the model's chat templates.
  • Test how the model handles indexed text sequences within its context window.

Example

import json, re
from pathlib import Path
import torch
from huggingface_hub import HfApi, hf_hub_download
from transformers import AutoTokenizer, AutoModelForImageTextToText

repo = "thurstonpowers/xll-acts-qwen3.5-9b"
revision = HfApi().model_info(repo, revision="acts-20261118").sha
system = 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 = 16384
if 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)
Read moreDetails
Benchmark on 200 acts (100 CT + 100 federal): 143/200 (71.5%) produced a structurally valid tree, 55 returned invalid output, and 2 hit the generation limit.
Tested on an RTX 5090 in BF16 with batch size 1, greedy decoding, and flash SDP (65,536-token context, up to 16,384 output tokens).
Trained using transformers 5.15.1 and PyTorch 2.13.0+cu130.
Released under the Apache-2.0 license.
Discuss Telegram 0 285

Will it run?

Runs in BF16

Full precision does not fit, a quantised version does: quality barely changes down to 8 bits, more noticeably at 4 bits.

  • F32weights 35 GB · ready files39 GB
  • BF16weights 18 GB20 GB
  • FP8 / INT8recommendedweights 8.8 GB10 GB
  • 4-bitweights 4.9 GB5.9 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2080 Ti

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # thurstonpowers/xll-acts-qwen3.5-9b
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download thurstonpowers/xll-acts-qwen3.5-9b --local-dir ~/models/xll-acts-qwen3.5-9b
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to xll-acts-qwen3.5-9b
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

  1. Qwen3.5-9B-Basefamily root
  2. Qwen3.5-9B
  3. xll-acts-qwen3.5-9b

New posts about versions of Qwen3.5-9B-Base (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 6All on Hugging Face
Merges 1All on Hugging Face

About the model

thurstonpowers/xll-acts-qwen3.5-9b
Author
thurstonpowers
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

2850

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments