Skip to content
Other

Qwen3-4B NACHA ACH writer

LoRA, которая собирает ACH-файлы байт в байт

1 min read

LoRA-адаптер для Qwen/Qwen3-4B, который превращает словесную платёжную спецификацию в полный NACHA ACH-файл. Сначала модель пишет worksheet с подсчётами и паддингом, затем сам файл; результат проверяется кодом на точное совпадение.

Ключевые особенности

  • На входе — originator, батчи, записи с routing numbers, счетами и суммами; на выходе — готовый ACH-файл.
  • Учитывает 94-символьные записи, padding, batch/file totals, entry hashes и 9-filler block count.
  • Сначала строит worksheet: нумерует записи, считает блоки, расписывает pad/cut по 5-символьным группам и running sums по 3-значным.
  • В репозитории есть eval-наборы, training script и проверка exact match кодом.
  • Заявлены 45/50 на eval и 99/100 на holdout со свежими seed'ами.

Пример

python eval.py --model paawandesai/qwen3-4b-nacha-writer --eval-set data/ach/holdout.jsonl --engine vllm
ПодробнееТочные файлы: eval 45/50, holdout 99/100; на первых 20 eval-спеках — 18/20 (Claude Opus 5.5 — 17/20, GPT-5.5 — 3/20, Claude Sonnet 5.5 — 1/20).
Обучение: LoRA r32/alpha64 на всех projections, lr 2e-4 cosine, batch 8, max length 32 768, bf16, TRL SFTTrainer, 1 эпоха на одной L40S; данные синтетические.
Ограничение: большие файлы (2–3 батча по 15–24 записи) ещё ошибаются в длинных running sums — чаще всего в нескольких строках control totals.
Discuss Telegram 1 62

Will it run?

Working out the memory needs

The bot is still collecting this model's weight sizes; the estimate will appear here in a few minutes. Meanwhile you can rent a GPU with headroom.

Massed ComputeOpen Massed Compute

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.

    # paawandesai/qwen3-4b-nacha-writer
    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 paawandesai/qwen3-4b-nacha-writer --local-dir ~/models/qwen3-4b-nacha-writer
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to qwen3-4b-nacha-writer
  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

New posts about versions of qwen3-4b-nacha-writer (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

paawandesai/qwen3-4b-nacha-writer
Author
paawandesai
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

621

Ask AI about this model

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

Comments

Other

homi

homi