Skip to content
Other

Code RL LR Ablation

где GRPO на MBPP теряет стабильность

1 min read

Это набор экспериментальных чекпоинтов от RL-Forgettting-Exp4, а не готовая продуктовая LLM. Абляция проверяет, как actor learning rate влияет на стабильность GRPO при обучении коду на MBPP: три значения ниже дефолтного 1e-6 — 6e-6, 3e-6 и 6e-7, каждое с replay-буфером и без. Базы — Qwen/Qwen3-1.7B-Base и Qwen/Qwen2.5-3B.

Ключевые особенности

  • GRPO на MBPP: лестница actor lr 6e-6, 3e-6 и 6e-7; каждый запуск повторяется с replay и без.
  • Базы: Qwen/Qwen3-1.7B-Base и Qwen/Qwen2.5-3B; в таблице карточки приведены три ячейки для lr 6e-7.
  • Replay-ветка: 128 прошлых rollouts на шаг, weight lambda=0.1, выбор через hard_cooldown.
  • Данные: 320 задач MBPP, вынесенных из MBPP+ (378) и канонического test split (276).
  • В каждом каталоге — полный HF-чекпоинт global_step_N, логи тренера и сэмплированные rollouts.

Пример

from transformers import AutoModelForCausalLM
m = AutoModelForCausalLM.from_pretrained(
    "RL-Forgettting-Exp4/code-rl-lr-ablation",
    subfolder="qwen3_1p7b_base_mbpp_nobuf_lr6e7/global_step_1000"
)
ПодробнееОценка: MBPP+ (378 задач), n=160 сэмплов, temperature 0.6, top_p 0.95, unbiased pass@k; полный тест-сьют MBPP+ (~108 тестов на задачу).
Тренировка: GRPO, 8 rollouts на промпт, batch 64, длина ответа 3072; бинарная награда за прохождение asserts.
В карточке нет итоговых цифр pass@k и требований к VRAM; выше 1e-6: 1e-5 — «coin flip», 1e-4/1e-3 стабильно фатальны для обеих веток.
Discuss Telegram 1

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 3.8 GB · ready files4.7 GB
  • FP8 / INT8recommendedweights 1.9 GB2.6 GB
  • 4-bitweights 1.1 GB1.7 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.

    # RL-Forgettting-Exp4/code-rl-lr-ablation
    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 RL-Forgettting-Exp4/code-rl-lr-ablation --local-dir ~/models/code-rl-lr-ablation
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to code-rl-lr-ablation
  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 code-rl-lr-ablation (quants, finetunes, LoRAs) will show up under the bell.

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

About the model

RL-Forgettting-Exp4/code-rl-lr-ablation
Author
RL-Forgettting-Exp4
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

01

Ask AI about this model

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

Comments