Skip to content
Other

lora-roberta-base-glue

LoRA adapters for RoBERTa-base on GLUE tasks

2 min readupdated

A set of weights and code to reproduce LoRA (a lightweight add-on on top of a base model for efficient fine-tuning) training across five GLUE benchmark tasks (SST-2, MRPC, CoLA, RTE, STS-B). Built on a frozen roberta-base (FacebookAI/roberta-base) with rank r = 8 and alpha = 8, the adapters target the query and value projections across all 12 attention layers alongside a trained classification head. This repository is useful for researchers and developers exploring practical LoRA applications on classic language models.

Key features

  • Five pre-trained weight sets for SST-2, MRPC, CoLA, RTE, and STS-B.
  • From-scratch LoRA implementation built on top of a frozen roberta-base model.
  • Each adapter includes a trained classification head.
  • Detailed evaluation results across 5 seeds for every task.

What you can do

  • Benchmark the pre-trained adapters on GLUE validation splits.
  • Use the repository code as a base for your own LoRA adaptation experiments.
  • Study how skipping the intermediate MNLI step affects adapter performance (for example, on RTE).

Example

import json
import torch
from huggingface_hub import snapshot_download
from safetensors.torch import load_file
from transformers import AutoModelForSequenceClassification, AutoTokenizer

task = "rte"
root = snapshot_download("Bartu0nder/lora-roberta-base-glue")
cfg = json.load(open(f"{root}/{task}/config.json"))

tok = AutoTokenizer.from_pretrained(cfg["base_model"])
model = AutoModelForSequenceClassification.from_pretrained(
    cfg["base_model"], num_labels=cfg["num_labels"], attn_implementation="sdpa")

state = load_file(f"{root}/{task}/adapter.safetensors")
print(f"Loaded adapter for {task} successfully")
Read moreDetails
Validation metrics (median across 5 seeds): SST-2 (Accuracy) — 94.8, MRPC (Accuracy) — 89.2, CoLA (Matthews corr.) — 63.1, RTE (Accuracy) — 78.7, STS-B (Pearson corr.) — 91.0.
Parameters: 294,912 trainable LoRA parameters and 592,130 classification head parameters.
Limitations: Adapters for MRPC, RTE, and STS-B were trained directly from roberta-base without an intermediate MNLI step, which accounts for score variations on RTE.
License: MIT.
Discuss Telegram 1

Will it run?

Runs in F32

Fits without quantisation.

  • F32weights 0.5 GB · ready files1.0 GB
  • BF16weights 0.2 GB0.8 GB
  • FP8 / INT8recommendedweights 0.1 GB0.6 GB
  • 4-bitweights 0.1 GB0.6 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 · F32 — from RTX 2060

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 roberta-base.

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.

    # Bartu0nder/lora-roberta-base-glue
    # ===== 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 FacebookAI/roberta-base --local-dir ~/models/roberta-base
    hf download Bartu0nder/lora-roberta-base-glue --local-dir ~/models/lora-roberta-base-glue
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to roberta-base
  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

FacebookAI/roberta-base

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

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

Quantised versions 7All on Hugging Face
Fine-tunes 12All on Hugging Face

About the model

Bartu0nder/lora-roberta-base-glue
Author
Bartu0nder
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