Skip to content
Language models

Prelude 9B SFT v2 Hallucination R2

an honest language model against hallucinations

2 min readupdated

A new research checkpoint (intermediate model version) with 9.09 billion parameters based on marianna13/prelude-9b-sft-v2 is trained to honestly admit when it does not know something instead of making up facts across ten European languages. The base model tends to invent biographies and plots for non-existent people or objects, whereas this version politely declines to answer and asks for context in 98% of cases. The model was created as part of OpenEuroLLM research and will be useful for developers who care about factual accuracy and minimizing fabrications in chat bots.

Key features

  • Sharp reduction in fabrications: on fictitious topics, the hallucination rate dropped from 97.7% to 1.8%.
  • Support for 10 European languages (English, German, French, Spanish, Italian, and others).
  • Multi-stage training using SFT (supervised fine-tuning) and DPO (direct preference optimization).
  • Retention of correct answers to known facts at 53.8% (though the model has become overly cautious).

What you can do

  • Test the model's resistance to fabrications (hallucinations) using your own tricky questions in different languages.
  • Integrate the model into text generation pipelines where refusing to answer in the absence of reliable data is critically important.
  • Use the standard Prelude ChatML chat template and tokenizer from the Qwen3 architecture for quick integration.
  • Explore the datasets and pipeline on GitHub to replicate experiments on combating hallucinations.

Example

from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "birgermoell/prelude-9b-sft-v2-hallucination-r2"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype="bfloat16", device_map="auto")
msgs = [{"role": "user", "content": "Wer ist Declan Monmonley?"}]
inputs = tok.apply_chat_template(msgs, add_generation_prompt=True, enable_thinking=False,
                                 return_tensors="pt", return_dict=True).to(model.device)
out = model.generate(**inputs, max_new_tokens=256)
print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
Read moreDetails
Benchmarks (500 questions): hallucinations on fictitious subjects decreased from 97.7% to 1.8%, and the truthfulness score increased from -85.2 to +3.8.
Requirements: two nodes with 4 × A100 (64 GB) for training, and inference requires a modern video card with bfloat16 support.
Limitations: the model has become too cautious and now frequently refuses to answer real questions about details of famous people (such as birth dates), and the thinking mode has lost factual accuracy.
Discuss Telegram 0 32

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 17 GB · ready files19 GB
  • FP8 / INT8recommendedweights 8.5 GB10 GB
  • 4-bitweights 4.8 GB6.5 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 · BF16 — from RTX 4090

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

Language models: with a 8192-token context.

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

Run it on Massed Compute

A ready script for the machine we advise above: paste it and the model downloads and starts.

  1. 1. Rent A30

    $0.26 an hour with the coupon. Pick the Ubuntu image when you create the machine.

    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.

    # birgermoell/prelude-9b-sft-v2-hallucination-r2 — A30, FP8 / INT8
    # NVIDIA driver: CUDA 13 needs 580+. If older or missing, the current nvidia-open is installed (keep yours: KEEP_DRIVER=1)
    DRV=$(nvidia-smi --query-gpu=driver_version --format=csv,noheader 2>/dev/null | head -1 | cut -d. -f1)
    case "$DRV" in ""|*[!0-9]*) DRV=0 ;; esac
    if [ "$DRV" -lt 580 ] && [ -z "${KEEP_DRIVER:-}" ]; then
      . /etc/os-release && \
      curl -fsSLO "https://developer.download.nvidia.com/compute/cuda/repos/ubuntu${VERSION_ID//./}/x86_64/cuda-keyring_1.1-1_all.deb" && \
      sudo dpkg -i cuda-keyring_1.1-1_all.deb && rm -f cuda-keyring_1.1-1_all.deb && \
      sudo apt-get update && \
      sudo DEBIAN_FRONTEND=noninteractive apt-get install -y nvidia-open || \
      { echo "Could not install the NVIDIA driver: see the error above. To keep the current one, run with KEEP_DRIVER=1."; exit 1; }
      echo "NVIDIA driver updated. Reboot the machine (sudo reboot) and run this same script again."
      exit 0
    fi
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv venv ~/vllm --python 3.12 && source ~/vllm/bin/activate
    uv pip install vllm
    vllm serve birgermoell/prelude-9b-sft-v2-hallucination-r2 --max-model-len 8192 --quantization fp8
  3. 3. Use it

    The server listens on port 8000 and speaks the OpenAI API: http://localhost:8000/v1 on the machine. From your computer: ssh -L 8000:localhost:8000 user@machine-IP.

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.

  • MTP (multi-token prediction)0.34.0from November 16, 2026new release

    1.5–2× generation speed

    If the model has MTP layers (DeepSeek-V3/R1, Qwen3-Next, GLM, MiMo and others), it predicts several tokens per step itself.

    How to enable: vLLM / SGLang: --speculative-config '{"method": "mtp", "num_speculative_tokens": 1}'. Check the model card.

    Earlier generations (2)
    • EAGLE-3 — Speculative decoding with a small draft head. For models without their own MTP layers.
    • EAGLE-2 / Medusa — Earlier draft-head generations. Still work if that is all you find.
  • FP8 KV cache0.34.0from November 16, 2026new release

    2× more context in the same memory

    Keeps the attention cache in 8 bits, so long contexts fit where they did not.

    How to enable: vLLM: --kv-cache-dtype fp8; llama.cpp: -ctk q8_0 -ctv q8_0.

  • vLLM / SGLang0.34.0from November 16, 2026new release

    many times more tokens/s on a server

    Server engines: batching, FlashAttention, FP8 and speculative decoding out of the box.

    How to enable: vllm serve <model> — an OpenAI-compatible API.

    Earlier generations (1)
    • llama.cpp — GGUF models on any GPU, Mac or even CPU; some layers can stay in RAM.

Model family

  1. prelude-9b-midtrain-v1family root
  2. prelude-9b-sft-v2
  3. prelude-9b-sft-v2-hallucination-r2

New posts about versions of prelude-9b-midtrain-v1 (quants, finetunes, LoRAs) will show up under the bell.

Fine-tunes 2All on Hugging Face

About the model

birgermoell/prelude-9b-sft-v2-hallucination-r2
Author
birgermoell
Category
Language models
Released
Posted

On NeyroLearn

Hugging Face

320

Ask AI about this model

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

Comments