This project summarizes long interview transcripts using the DeepSeek LLaMA model (deepseek-ai/DeepSeek-R1-Distill-Llama-8B).

git clone https://github.com/Aditya3107/ATRIUM_summarization.git
cd ATRIUM_summarization

We recommend using a virtual environment:

python3 -m venv venv
source venv/bin/activate
pip install .

Export your token so the script can access private models:

Edit run.sh and replace:

--hf_token <YOUR HUGGINGFACE TOKEN>

with your actual token.

Place your .srt or speaker-labeled .txt transcript file into the inputs/ folder.

Example:

inputs/sample_interview.txt

Run the summarizer on one of the files in inputs, specify only the filename:

./run.sh sample_interview.txt

Summaries will be saved in the output/ folder with the extension summary.txt, for example:

sample_interview.summary.txt

This Docker image wraps a Python-based summarization pipeline using the DeepSeek LLaMA model and is optimized for GPU usage.

• GPU-enabled (NVIDIA CUDA 12.3)
• Accepts custom .srt or .txt files for summarization
• Mountable input/output and cache directories
• Hugging Face token support via environment variable

You can use our pre-built image hosted on Docker Hub:

docker pull aditya3107/atrium-summarizer:latest

docker run --rm \
  --gpus all \
  -e HF_TOKEN=your_actual_token_here \
  -v $(pwd)/inputs:/app/inputs \
  -v $(pwd)/output:/app/output \
  -v $(pwd)/cache:/app/cache \
  summarizer-gpu \
  --srt-file /app/inputs/sample_data2.txt \
  --intro-prompt "Jonathan Carker interviewing Cheryl Jones on 30th September at Grand Union's magnificent Bothy." \
  --model-name deepseek-ai/DeepSeek-R1-Distill-Llama-8B \
  --summary-words 1000 \
  --use-gpu yes \
  --device-id 0 \
  --cache-dir /app/cache \
  --hf-token $HF_TOKEN