Skip to content
Other

Beneficial trait RL

LoRA для полезных черт gpt-oss-20b

2 min readupdated

Это LoRA-адаптеры (небольшие надстройки над базовой моделью) для openai/gpt-oss-20b, обученные в рамках открытой репликации статьи о reinforcement learning для «широко и устойчиво полезных» моделей. Авторы сравнили два прогона: обычный RL на медицинских и математических задачах и trait RL, где 25% каждого батча заменяли синтетическими health-диалогами с наградой за truthfulness, corrigibility, metacognitive transparency и другие полезные черты. На выходе — rank-32 LoRA в формате PEFT для attention, MLP, expert-слоёв и output head, с чекпоинтами на 20, 50, 80 и 100 шагах. Пригодится исследователям, которые хотят проверить или использовать адаптер с улучшенными «полезными» качествами; это исследовательские чекпоинты, не для медицинских советов.

Ключевые особенности

  • Два набора чекпоинтов: normal_rl и beneficial_trait_rl, по шагам 20/50/80/100.
  • Rank-32 LoRA (lora_alpha 32) в PEFT для attention, MLP, expert-слоёв и output head.
  • Обучение на Tinker: 100 шагов, 32 промпта × 16 сэмплов, learning rate 4e-5, Harmony chat format и medium reasoning effort.
  • На step 100 trait RL обгоняет normal RL на held-out trait benchmark на +21.5 пункта (95% CI [17.2, 25.8]).
  • По карточке, нет значимой потери на MedMCQA, MedQA, MedCalc-Bench, MATH, GPQA Diamond, HMMT и IFEval.

Что можно сделать

  • Подключить step_100 через vLLM как LoRA-модуль trait_rl и сравнить ответы с базовой gpt-oss-20b.
  • Загрузить адаптер через PEFT и исследовать, как меняются truthfulness, corrigibility и metacognitive transparency в диалогах.
  • Использовать normal_rl как compute-matched baseline в собственных экспериментах с RL.
  • Тестировать health-сценарии, помня, что чекпоинты не предназначены для медицинских советов.

Пример

hf download Aansh123/beneficial-trait-rl --include "beneficial_trait_rl/step_100/*" --local-dir ckpt
vllm serve openai/gpt-oss-20b --enable-lora --max-lora-rank 32 \
  --lora-modules trait_rl=ckpt/beneficial_trait_rl/step_100
ПодробнееБенчмарки step 100 (trait RL минус normal RL, 95% CI): held-out trait +21.5 [17.2, 25.8]; MASK +26.8 [14.6, 39.0]; DeceptionBench +15.0 [9.4, 21.2].
HealthBench +13.1 [7.1, 19.0]; sycophancy +10.1 [4.3, 16.3].
Обучение: Tinker, 100 шагов, один seed на запуск; чекпоинты исследовательские, не для медицинских советов.
Discuss Telegram 0

Will it run?

Runs in FP8 / INT8

Full precision does not fit, a quantised version does: quality barely changes down to 8 bits, more noticeably at 4 bits.

  • U8weights 23 GB · ready files26 GB
  • FP8 / INT8recommendedweights 21 GB24 GB
  • 4-bitweights 13 GB14 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 4090 · U8 — from RTX 5090

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 gpt-oss-20b.

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.

    # Aansh123/beneficial-trait-rl
    # ===== 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 openai/gpt-oss-20b --local-dir ~/models/gpt-oss-20b
    hf download Aansh123/beneficial-trait-rl --local-dir ~/models/beneficial-trait-rl
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to gpt-oss-20b
  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

openai/gpt-oss-20b

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

New posts about versions of gpt-oss-20b (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 2All on Hugging Face
Merges 2All on Hugging Face

About the model

Aansh123/beneficial-trait-rl
Author
Aansh123
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

Other

CRISP

CRISP