Skip to content
Other

Qwen3.8-27B-BarronPolyStep-RealCoder-v3

PEFT LoRA для Python-кода с аудитом zero-leakage

2 min read

Это LoRA-адаптер для Qwen/Qwen3.8-27B от ProbioticFarmer: 64 слоя, 256 тензоров, r=16, lora_alpha=32, target modules down_proj, o_proj, out_proj. Обучение шло на 40 задачах (160 слотов, 3 378 token positions) на Apple M4 Max 128 GB; 16 тестовых задач (64 слота, 1 327 token positions) держали полностью вне обучения.

Ключевые особенности

  • Строгий zero-leakage аудит: проверка audit_realcoder_v3.py подтверждает стандартный peft.LoraConfig, отдельные подпространства для каждого из 64 слоёв и max_abs_diff = 0.0 при замене held-out позиций шумом.
  • Barron-Patched PolyStep: BarronCrossEntropyLoss + barron_barycentric_centroid, обучение на 40-task split.
  • Бенчмарки: combined Pass@1 48/56 (85.71%) у Barron Full-Fixed против 36/56 (64.29%) у базовой модели в протоколе 40 train / 16 held-out.
  • Публичные лидерборды: HumanEval+ 137/164 (83.54%) против 117/164 (71.34%), MBPP+ 297/378 (78.57%) против 281/378 (74.34%), BigCodeBench-Hard 45/148 (30.41%) против 41/148 (27.70%).
  • Ограничение: обучение и валидация — на небольшом наборе Python-задач; заявленные результаты стоит проверять на своих данных.

Пример

Сценарий из карточки: генерация Python-кода для задач Algorithm & Systems и Python & Data Science с проверкой через compile() + exec(). Для open-ended генерации в карточке указан lora_scale=2.0 и max_new_tokens=104.

ПодробнееОбучение: 40 задач / 160 слотов / 3 378 token positions на Apple M4 Max 128 GB; held-out — 16 задач / 64 слота / 1 327 token positions.
Оценка: NVIDIA A100-SXM4-80GB; HumanEval+ 83.54%, MBPP+ 78.57%, BigCodeBench-Hard 30.41% у RealCoder-v3.
Ограничения: маленький train split, отчёт основан на собственных замерах автора; адаптер требует базовую Qwen/Qwen3.8-27B и peft.
Discuss Telegram 0 372

Will it run?

Runs in 4-bit

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

  • BF16weights 52 GB · ready files57 GB
  • FP8 / INT8recommendedweights 26 GB29 GB
  • 4-bitweights 15 GB17 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.

This is a LoRA: counted together with its base model Qwen3.8-27B.

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.

    # ProbioticFarmer/Qwen3.8-27B-BarronPolyStep-RealCoder-v3
    # ===== 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 Qwen/Qwen3.8-27B --local-dir ~/models/Qwen3.8-27B
    hf download ProbioticFarmer/Qwen3.8-27B-BarronPolyStep-RealCoder-v3 --local-dir ~/models/Qwen3.8-27B-BarronPolyStep-RealCoder-v3
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Qwen3.8-27B
  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

Qwen/Qwen3.8-27B

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

New posts about versions of Qwen3.8-27B (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

About the model

ProbioticFarmer/Qwen3.8-27B-BarronPolyStep-RealCoder-v3
Author
ProbioticFarmer
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

3720

Ask AI about this model

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

Comments

Language models

joey-1

joey-1