Skip to content
Computer vision

tc-4b

Torchcast Decision 4B, continued: прирост Capability на JevBench

2 min read

tc-4b — продолжение обучения torchcast-ai/torchcast-decision-4b v1.3.1 (база Qwen3.5-4B) от yujie111222. Это LoRA-дообучение (r=32, alpha=64, lr 1e-4, 49 optimizer steps) с последующим merge весов на ~94k admitted rows из лицензионно отфильтрованной стратифицированной выборки DecisionBank. Размер — около 5,17B параметров; лицензия Apache 2.0.

Ключевые особенности

  • LoRA r=32, alpha=64, lr 1e-4, 49 optimizer steps; веса объединены с базовой моделью.
  • Обучение на ~94k admitted rows из DecisionBank-260k; использован лицензионный allowlist apache/mit/bsd/cc-by/cc0/unlicense/public-domain, public-split JevBench не применялся.
  • На JevBench 231 public items (v1.6 scorer): Capability 85.54 против 81.37 у базы, Intelligence 79.81 против 74.93, Calibration 91.26 против 87.82.
  • Сервинг как у базовой модели: bundled torchcast_decision/ и decision_config.json; full-attention-слои noncausal.
  • vLLM и llama.cpp не воспроизводят оценённый интерфейс из-за generic causal-attention.
  • Apache 2.0; модель производна от torchcast-decision-4b и Qwen3.5-4B.

Пример

Сценарий из карточки — decision-задачи, где важны качество распределения и expected-value distance, а не только argmax accuracy. Готового промпта в карточке нет: запускать стоит через bundled torchcast_decision/ и decision_config.json, как базовую v1.3.1 — обычный causal-инференс не гарантирует совпадения с оценённым интерфейсом.

ПодробнееJevBench 231 public items, v1.6 scorer: Intelligence 79.81, Calibration 91.26, Capability 85.54; у базы v1.3.1 — 74.93 / 87.82 / 81.37.
Прирост сосредоточен в score question type: качество распределения и expected-value distance, а не argmax accuracy.
Требований к VRAM в карточке нет; для оценки нужен bundled serving, generic causal-рантаймы не подходят.
Данные: ~94k admitted rows; JevBench public-split не использовался.
Discuss Telegram 0 26

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.

    # yujie111222/tc-4b
    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 yujie111222/tc-4b --local-dir ~/models/tc-4b
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to tc-4b
  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 tc-4b (quants, finetunes, LoRAs) will show up under the bell.

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

About the model

yujie111222/tc-4b
Author
yujie111222
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

260

Ask AI about this model

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

Comments