Skip to content
Other

CLM-8B reproduction heads

независимое воспроизведение CLM для Qwen3-8B

2 min readupdated

Это независимое воспроизведение CLM-8B («Contrastive Language Models», Kwok et al., Stanford и NVIDIA Research, 2026) от JaydeepR: две небольшие projection-головы поверх замороженной Qwen3-8B. Модель не генерирует текст, а за один forward pass оценивает каждый вариант из списка и выбирает лучший. Суммарно обучено 18,9 млн параметров.

Ключевые особенности

  • Две головы — State head и Action head — поверх замороженной Qwen3-8B; архитектура каждой: 4096 → 1536 → 1536 (LayerNorm) → 512 с GELU.
  • На входе ситуация и список опций; на выходе — softmax по косинусному сходству с обучаемым масштабом. Текст не генерируется.
  • Чекпоинты в формате CLM: state_head, action_head, logit_scale, cfg; загружаются в сервере авторов.
  • На Tricky Q&A после post-training top-1 69,0% — чуть выше 68,6% у голов авторов; на general Q&A 82,7% против 83,6%.
  • На game-style T-Rex заметно хуже: agreement с планировщиком игры 2,7% против 65,3%, safety shield 3/5 против 5/5.

Пример

pip install git+https://github.com/Contrastive-LM/CLM.git
vllm serve Qwen/Qwen3-8B --runner pooling --max-model-len 2048 --port 8090
clm-serve --ckpt clm-8b-repro-post.pt
ПодробнееTricky Q&A post-training top-1: 69,0% (published 68,5%, heads авторов 68,6%); general Q&A 1,024-way: 82,7% (83,6%).
DeepSWE best-of-4, среднее по 5 seed: 30,8/38 (31,0/38); медианная задержка решения на L40S — 17,1 ms.
Обучение на Modal, около $82: pre — 30 000 шагов, mid — 1 344, post — 400, batch 16 384.
Нужен encoder Qwen3-8B revision b968826d9c46dd6066d109eabc6255188de91218, last-token pooling и L2-нормализация.
Ограничения: T-Rex сильно хуже голов авторов, ручные промпты ненадёжны (1 из 8), только английский текст.
Discuss Telegram 0

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.1 GB · ready files0.6 GB
  • FP8 / INT8recommendedweights 0.0 GB0.5 GB
  • 4-bitweights 0.0 GB0.5 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 2060 · BF16 — from RTX 2060

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.

    # JaydeepR/clm-8b-repro-heads
    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 JaydeepR/clm-8b-repro-heads --local-dir ~/models/clm-8b-repro-heads
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to clm-8b-repro-heads
  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. Qwen3-8B-Basefamily root
  2. Qwen3-8B
  3. clm-8b-repro-heads

New posts about versions of Qwen3-8B-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 4All on Hugging Face

About the model

JaydeepR/clm-8b-repro-heads
Author
JaydeepR
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