Skip to content
Other

TwIL-8B-pi

LoRA для агента pi в анализе данных

1 min readupdated

LoRA-адаптер от harness-hub для базовой модели webAI-Official/TwIL-8B (8B). Обучен с подкреплением под pi coding agent: агент читает задачу, запускает shell и Python, отвечает, а верификатор оценивает результат.

Ключевые особенности

  • Обучение через AsyncGRPO из TRL: LoRA r=32, alpha=64, все линейные слои.
  • На 150 held-out задачах SmolDataEnvs (n=146) — 84.9% против 81.5% у базы (+3.4 ± 5.8 п.п.).
  • На задачах, которые решают обе модели, на 37% меньше вызовов инструментов (112 задач).
  • Заточен под harness pi; для vLLM есть конфиг с LoRA rank 32 и парсерами qwen3_coder/qwen3.
  • Некоммерческая лицензия webAI; вне pi ожидайте поведение базовой модели или хуже.

Пример

vllm serve webAI-Official/TwIL-8B --enable-lora --max-lora-rank 32 --lora-modules twil-harbor-pi=harness-hub/TwIL-8B-pi --enable-auto-tool-choice --tool-call-parser qwen3_coder --reasoning-parser qwen3 --chat-template chat_template.jinja --max-model-len 131072
ПодробнееБенчмарк: 84.9% vs 81.5% на 150 held-out задачах SmolDataEnvs (+3.4 ± 5.8 п.п.); на 40 задачах traj2envs (n=35) — 37.1% vs 31.4% (+5.7 ± 16.0).
Эффективность: −37% tool calls на 112 общих решённых задачах; шаг 90 из 100 выбран по той же eval, поэтому прирост оптимистичен.
Ограничения: малая eval (±5.8 п.п.), только анализ данных и harness pi; GGUF-версия не переоценена; лицензия некоммерческая.
Обучение: 4× NVIDIA H200, 8 rollout на задачу, до 30 вызовов модели и 1200 с на rollout.
Discuss Telegram 1 8

Will it run?

Runs in GGUF F16

The most precise of the ready quantised files fits.

  • GGUF F16weights 17 GB · ready files19 GB
  • BF16weights 17 GB · ready files19 GB
  • GGUF Q4_K_Mweights 10 GB · ready files12 GB
  • GGUF Q8_0recommendedweights 9.1 GB · ready files10 GB
  • FP8 / INT8weights 8.6 GB9.9 GB
  • 4-bitweights 5.0 GB6.0 GB

Tip: go 8-bit.GGUF Q8_0 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: GGUF Q8_0 — from RTX 2080 Ti · GGUF F16 — from RTX 4090

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 TwIL-8B.

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.

    # harness-hub/TwIL-8B-pi
    # ===== 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 webAI-Official/TwIL-8B --local-dir ~/models/TwIL-8B
    hf download harness-hub/TwIL-8B-pi --local-dir ~/models/TwIL-8B-pi
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to TwIL-8B
  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

webAI-Official/TwIL-8B

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

  1. granite-4.1-8b-basefamily root
  2. granite-4.2-8b
  3. TwIL-8B
  4. TwIL-8B-pi

New posts about versions of granite-4.1-8b-base (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 3All on Hugging Face
Fine-tunes 6All on Hugging Face

About the model

harness-hub/TwIL-8B-pi
Author
harness-hub
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

81

Ask AI about this model

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

Comments