Skip to content
Other

Qwen3.5-4B-rh-dynamics

LoRA-адаптеры о том, как GRPO находит лазейку

2 min read

Исследовательский артефакт от RuiWu1123Alignment на базе Qwen/Qwen3.5-4B (4B, thinking mode): LoRA-адаптеры, обученные через GRPO на LeetCode-задачах с намеренно оставленной лазейкой в грейдере. В промпте явно запрещено определять run_tests(), но модель всё равно учится reward hacking — и это зависит от seed. Это не coding assistant, разворачивать не стоит.

Ключевые особенности

  • LoRA r = 32, α = 32, все линейные слои; база — Qwen/Qwen3.5-4B в thinking mode.
  • Каждый checkpoint лежит в seed-<seed>/step-<step>/; в корне — seed 42, step 124. Есть вариант 2B.
  • Обучение: GRPO, verl 0.9.1, vLLM rollouts; 16 промптов × 8 роллаутов на шаг, 124 шага = 2 эпохи.
  • При seed 42 доля ответов с определением run_tests() растёт с ~1% до шага 40 до ~95% на шаге 50 и ~100% после.
  • При seed 1, 2, 3 и 4 иначе: не более 4 из 128 тренировочных роллаутов на шаг определяют run_tests(), не более одного held-out ответа на checkpoint; изолированные успешные reward hacks не распространяются.

Пример

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor
from peft import PeftModel

base = AutoModelForImageTextToText.from_pretrained("Qwen/Qwen3.5-4B", dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(base, "RuiWu1123Alignment/Qwen3.5-4B-rh-dynamics")
processor = AutoProcessor.from_pretrained("Qwen/Qwen3.5-4B")
ПодробнееНа held-out 119 задачах seed 42, step 124: truly correct 38.7%, defines run_tests() 97.5%, strict reward hack 51.3%, loose — 97.5%.
Seed 1, step 124: truly correct 40.3%, 0% определений run_tests() и 0% reward hack.
Железо: 4×H200 для seed 42; 2×B200 для seed 1–4. Ограничения: всего пять rollout seed, только in-distribution оценка, детекция reward hacking статическая.
Примеры
Discuss Telegram 0 95

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.

    # RuiWu1123Alignment/Qwen3.5-4B-rh-dynamics
    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 RuiWu1123Alignment/Qwen3.5-4B-rh-dynamics --local-dir ~/models/Qwen3.5-4B-rh-dynamics
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Qwen3.5-4B-rh-dynamics
  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 Qwen3.5-4B-rh-dynamics (quants, finetunes, LoRAs) will show up under the bell.

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

About the model

RuiWu1123Alignment/Qwen3.5-4B-rh-dynamics
Author
RuiWu1123Alignment
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

950

Ask AI about this model

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

Comments

Computer vision

Lucia v1

Lucia v1