Skip to content
Other

giocorte/totem-kiosk-extractor-sft-v2: LoRA adapters for ticket kiosks built on Qwen3

2 min read

This release introduces two new LoRA adapters (lightweight fine-tunes built on top of a base model) trained jointly on an expanded dataset for an automated ticketing kiosk powered by Qwen/Qwen3-4B-Instruct-2507. The first adapter manages customer dialogue through a command-line interface, while the second handles precise order-state extraction based on each incoming message. The models are trained to follow specifications strictly: using exact tool names, correctly handling locations, and warning users when default values are applied. This solution is well-suited for developers building voice assistants, terminal chatbots, and automated customer service systems.

Key features

  • Dual adapter: A linked setup featuring the kiosk_v2 module for dialogue and the extractor_v2 module for order parameter extraction.
  • High tool-name accuracy: Exact matches for tool names increased to 86% thanks to a refined dataset.
  • SMS logic: Requests to send info to a phone now correctly map to SMS instead of email in 14 out of 15 cases.
  • Contextual zones: The phrase "from here" automatically binds to the kiosk's current location.

What you can do

  • Deploy kiosk backends: Run the adapter bundle via vLLM to power real voice or text terminals.
  • Test dialogue scenarios: Use the built-in set of 50 realistic dialogues with language agents to evaluate performance quality.
  • Integrate into interfaces: Configure a dialogue agent for purchasing train, bus, or ferry tickets across multiple languages.

Example

An example of running two LoRA models using the vLLM server:

vllm serve Qwen/Qwen3-4B-Instruct-2507 --max-model-len 32768 --enable-lora --max-lora-rank 32 --max-loras 2 \
  --lora-modules kiosk_v2=./kiosk_v2 extractor_v2=./extractor_v2
Read moreDetails
• Dataset: 460 examples for the kiosk and 3,146 for state extraction
• Training: 3 epochs for the kiosk, 2 epochs for extraction on an H200 NVL GPU
• Benchmark results: Out of 28 simulated customer flows, successful purchases were completed in 23 cases
• Limitations: Errors in processing spoken ambiguous time instructions and relative dates
Discuss Telegram 0

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 7.7 GB · ready files9.0 GB
  • FP8 / INT8recommendedweights 4.0 GB4.9 GB
  • 4-bitweights 2.4 GB3.1 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 2060 · BF16 — from RTX 3080

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

This is a LoRA: counted together with its base model Qwen3-4B-Instruct-2507.

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.

    # giocorte/totem-kiosk-extractor-sft-v2
    # ===== Tokens: paste yours between the quotes =====
    export HF_TOKEN=""        # huggingface.co/settings/tokens, for gated models
    # ==========================================
    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 Qwen/Qwen3-4B-Instruct-2507 --local-dir ~/models/Qwen3-4B-Instruct-2507
    hf download giocorte/totem-kiosk-extractor-sft-v2 --local-dir ~/models/totem-kiosk-extractor-sft-v2
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Qwen3-4B-Instruct-2507
  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

Requires the base model

Qwen/Qwen3-4B-Instruct-2507

This model is a LoRA / adapter. Download the base model first, then load this one on top.

New posts about versions of Qwen3-4B-Instruct-2507 (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 4All on Hugging Face

About the model

giocorte/totem-kiosk-extractor-sft-v2
Author
giocorte
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

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

Comments

Language models

Totem 4B

Totem 4B