Skip to content
Language models

T5Gemma 2B-2B · YouTube transcript chaptering

LoRA for splitting videos into chapters

2 min readupdated

This is a LoRA (a lightweight adapter built on top of a base model) for google/t5gemma-2b-2b-ul2-it by Saltywan. It automatically structures long-form videos: you pass in a YouTube transcript split into numbered sentences, and the model outputs a strict JSON table of contents specifying the starting sentence number and title for each chapter. Built as part of the llm-video-chaptering project, this adapter became the top-performing model in its series, outperforming both the base model and the specialized segmenter MiniSeg. It is a solid fit for developers building content-processing tools and NLP researchers.

Key features

  • High segmentation accuracy: Boundary F1 jumped from 0.149 on the base model to 0.500, reaching 0.679 with a two-sentence tolerance window.
  • Reliable JSON output: syntax errors during generation dropped from 477 down to just 4 failures across 1,448 test videos.
  • Handles long transcripts: the base context window fits up to 8,192 tokens, while longer transcripts are split into overlapping chunks and stitched back together.
  • Ready-to-use tooling: the project's GitHub repository includes scripts for prompt formatting, window chunking, and strict output parsing.

What you can do

  • Plug it into an English podcast or livestream processing pipeline to automatically generate timestamps from subtitles.
  • Build a bot or web app that takes a YouTube lecture transcript and returns a clean, sentence-indexed table of contents.
  • Use the code in the llm-video-chaptering repository to benchmark open models on text segmentation tasks.

Example

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
from peft import PeftModel

base = "google/t5gemma-2b-2b-ul2-it"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForSeq2SeqLM.from_pretrained(base, torch_dtype="bfloat16", device_map="cuda")
model = PeftModel.from_pretrained(model, "Saltywan/t5gemma-2b-2b-ytseg-chaptering-lora").eval()
# Полный пайплайн с build_window_prompt и parse_window_outline доступен в репозитории автора
Read moreDetails
Benchmarks (YTSeg, 1,448 videos): Boundary F1 exact 0.500 (vs. 0.149 base), Boundary F1 micro 51.99 (vs. 45.81 for MiniSeg), Title BERTScore 0.651.
Hardware and speed: inference on windows up to 8k tokens takes about 13 GiB of VRAM in bf16; training took 29 minutes on a single NVIDIA B200.
Training details: LoRA rank 16, alpha 32, dropout 0.05 across encoder, decoder, and cross-attention projections (adapter size is 73 MB).
Limitations: English only, text-only analysis with no audio or video input; licensed for non-commercial research only (CC BY-NC-SA 4.0).
Discuss Telegram 0 28

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 10 GB · ready files12 GB
  • FP8 / INT8recommendedweights 5.3 GB6.4 GB
  • 4-bitweights 3.0 GB3.9 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 5060 Ti 8GB · BF16 — from RTX 5080

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 t5gemma-2b-2b-ul2-it.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

A ready script for the machine we advise above: paste it and the model downloads and starts.

  1. 1. Rent A30

    $0.26 an hour with the coupon. Pick the Ubuntu image when you create the machine.

    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.

    # Saltywan/t5gemma-2b-2b-ytseg-chaptering-lora — A30, FP8 / INT8
    # ===== Tokens: paste yours between the quotes =====
    export HF_TOKEN=""        # huggingface.co/settings/tokens, for gated models
    # ==========================================
    # NVIDIA driver: CUDA 13 needs 580+. If older or missing, the current nvidia-open is installed (keep yours: KEEP_DRIVER=1)
    DRV=$(nvidia-smi --query-gpu=driver_version --format=csv,noheader 2>/dev/null | head -1 | cut -d. -f1)
    case "$DRV" in ""|*[!0-9]*) DRV=0 ;; esac
    if [ "$DRV" -lt 580 ] && [ -z "${KEEP_DRIVER:-}" ]; then
      . /etc/os-release && \
      curl -fsSLO "https://developer.download.nvidia.com/compute/cuda/repos/ubuntu${VERSION_ID//./}/x86_64/cuda-keyring_1.1-1_all.deb" && \
      sudo dpkg -i cuda-keyring_1.1-1_all.deb && rm -f cuda-keyring_1.1-1_all.deb && \
      sudo apt-get update && \
      sudo DEBIAN_FRONTEND=noninteractive apt-get install -y nvidia-open || \
      { echo "Could not install the NVIDIA driver: see the error above. To keep the current one, run with KEEP_DRIVER=1."; exit 1; }
      echo "NVIDIA driver updated. Reboot the machine (sudo reboot) and run this same script again."
      exit 0
    fi
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv venv ~/vllm --python 3.12 && source ~/vllm/bin/activate
    uv pip install vllm
    vllm serve google/t5gemma-2b-2b-ul2-it --max-model-len 8192 --quantization fp8 --enable-lora --max-lora-rank 64 --lora-modules t5gemma-2b-2b-ytseg-chaptering-lora=Saltywan/t5gemma-2b-2b-ytseg-chaptering-lora
    Add more LoRAs to t5gemma-2b-2b-ul2-it
  3. 3. Use it

    The server listens on port 8000 and speaks the OpenAI API: http://localhost:8000/v1 on the machine. From your computer: ssh -L 8000:localhost:8000 user@machine-IP.

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.

  • MTP (multi-token prediction)0.34.0from November 16, 2026new release

    1.5–2× generation speed

    If the model has MTP layers (DeepSeek-V3/R1, Qwen3-Next, GLM, MiMo and others), it predicts several tokens per step itself.

    How to enable: vLLM / SGLang: --speculative-config '{"method": "mtp", "num_speculative_tokens": 1}'. Check the model card.

    Earlier generations (2)
    • EAGLE-3 — Speculative decoding with a small draft head. For models without their own MTP layers.
    • EAGLE-2 / Medusa — Earlier draft-head generations. Still work if that is all you find.
  • FP8 KV cache0.34.0from November 16, 2026new release

    2× more context in the same memory

    Keeps the attention cache in 8 bits, so long contexts fit where they did not.

    How to enable: vLLM: --kv-cache-dtype fp8; llama.cpp: -ctk q8_0 -ctv q8_0.

  • vLLM / SGLang0.34.0from November 16, 2026new release

    many times more tokens/s on a server

    Server engines: batching, FlashAttention, FP8 and speculative decoding out of the box.

    How to enable: vllm serve <model> — an OpenAI-compatible API.

    Earlier generations (1)
    • llama.cpp — GGUF models on any GPU, Mac or even CPU; some layers can stay in RAM.

Model family

Requires the base model

google/t5gemma-2b-2b-ul2-it

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

  1. t5gemma-2b-2b-ul2family root
  2. t5gemma-2b-2b-ul2-it
  3. t5gemma-2b-2b-ytseg-chaptering-lora

New posts about versions of t5gemma-2b-2b-ul2 (quants, finetunes, LoRAs) will show up under the bell.

Fine-tunes 5All on Hugging Face

About the model

Saltywan/t5gemma-2b-2b-ytseg-chaptering-lora
Author
Saltywan
Category
Language models
Released
Posted

On NeyroLearn

Hugging Face

280

Ask AI about this model

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

Comments