Hearing Is Believing: Experiments & analysis code

In this repository, all code for experiments and analysis can be found to reproduce results from:

Schneider, F., & Blank, H. (2025). Sensory sharpening and semantic prediction errors unify competing models of predictive processing in in human speech comprehension. bioRxiv. 10.1101/2025.07.23.666354

🚀 Getting started

❗️ Required: Installation (Python)

Make sure you have anaconda installed on your system. Open your terminal, navigate to your root directory for this project and create a fresh environment, however suits you best:

1: On unix

On unix systems, we can use mamba for a quick installation like so:

brew install mamba
mamba env create -f environment.yml
conda activate sempriors

2: On Windows

On Windows, it'll be easiest to skip mamba and install from environment like so:

conda env create -f environment.yml
conda activate sempriors

💡 Hint: Using Intel Machines

If you are on an intel machine, please also run:

pip install intel-numpy
pip install scikit-learn-intelex

💡 Hint: PyTensor errors

Should you encounter PyTensor errors related to gcc/gxx, the easiest fix will be to run

pip install --upgrade pymc

❕ Optional: GPU Acceleration (Python)

Almost all scripts will allow you to specify GPU(s) to accelerate computations. If you would like to make use of this, please make sure to install an appropriate version of PyTorch that is compatible with your GPU.

For more information, please consult their offical documentation.

❗️ Required: Installation (R)

Please make sure to install R and, ideally, RStudio. Note that our machines run on R version 4.0.3 (2020-10-10). Once installed, please open RStudio and install the requirements like so:

install.packages(c('lme4', 'lmerTest', 'emmeans', 'DHARMa', 'ggplot2', 'viridis'))

❗️ Required: Downloading data

In your terminal, navigate to the top level of your working directory and download the data like so:

pip install osfclient
osf init
osf -p snxqm fetch -r / ./

Upon the init command, you may be prompted to input your OSF account and the project id, which is snxqm. Downloading all data may take a while, as the project is about 50GB. Please verify you have enough space before trying to download the data.

Alternatively, use your browser to navigate to https://osf.io/snxqm and download the full zip. Make sure to extract it to the top-level directory.

Once you have obtained the data, please navigate to your root directory and unzip individual folders like so:

1: On unix:

unzip ./analysis/eeg/data/raw/raw0.zip ./analysis/eeg/data/raw/
unzip ./analysis/eeg/data/raw/eeg/raw1.zip ./analysis/eeg/data/raw/eeg/
unzip ./analysis/eeg/data/raw/eeg/raw2.zip ./analysis/eeg/data/raw/eeg/
unzip ./analysis/eeg/data/raw/eeg/raw3.zip ./analysis/eeg/data/raw/eeg/
unzip ./analysis/eeg/data/raw/eeg/raw4.zip ./analysis/eeg/data/raw/eeg/
unzip ./analysis/eeg/data/raw/eeg/raw5.zip ./analysis/eeg/data/raw/eeg/
unzip ./analysis/eeg/data/preprocessed/audio/preprocessed0.zip ./analysis/eeg/data/preprocessed/audio/
unzip ./analysis/eeg/data/preprocessed/audio/w2v2/preprocessed1.zip ./analysis/eeg/data/preprocessed/audio/w2v2/
unzip ./analysis/eeg/data/preprocessed/preprocessed2.zip ./analysis/eeg/data/preprocessed/
unzip ./analysis/eeg/data/preprocessed/eeg/preprocessed3.zip ./analysis/eeg/data/preprocessed/eeg/
unzip ./analysis/eeg/data/preprocessed/eeg/preprocessed4.zip ./analysis/eeg/data/preprocessed/eeg/
unzip ./analysis/eeg/data/preprocessed/eeg/preprocessed5.zip ./analysis/eeg/data/preprocessed/eeg/
unzip ./analysis/eeg/data/processed/eeg/processed1.zip ./analysis/eeg/data/processed/eeg/
unzip ./analysis/eeg/data/processed/eeg/processed2.zip ./analysis/eeg/data/processed/eeg/
unzip ./analysis/eeg/data/results/results0.zip ./analysis/eeg/data/results/
unzip ./analysis/spaces/spaces.zip ./analysis/spaces/

2: On Windows:

tar -xzf ./analysis/eeg/data/raw/raw0.zip -C ./analysis/eeg/data/raw/
tar -xzf ./analysis/eeg/data/raw/eeg/raw1.zip -C ./analysis/eeg/data/raw/eeg/
tar -xzf ./analysis/eeg/data/raw/eeg/raw2.zip -C ./analysis/eeg/data/raw/eeg/
tar -xzf ./analysis/eeg/data/raw/eeg/raw3.zip -C ./analysis/eeg/data/raw/eeg/
tar -xzf ./analysis/eeg/data/raw/eeg/raw4.zip -C ./analysis/eeg/data/raw/eeg/
tar -xzf ./analysis/eeg/data/raw/eeg/raw5.zip -C ./analysis/eeg/data/raw/eeg/
tar -xzf ./analysis/eeg/data/preprocessed/audio/preprocessed0.zip -C ./analysis/eeg/data/preprocessed/audio/
tar -xzf ./analysis/eeg/data/preprocessed/audio/w2v2/preprocessed1.zip -C ./analysis/eeg/data/preprocessed/audio/w2v2/
tar -xzf ./analysis/eeg/data/preprocessed/preprocessed2.zip -C ./analysis/eeg/data/preprocessed/
tar -xzf ./analysis/eeg/data/preprocessed/eeg/preprocessed3.zip -C ./analysis/eeg/data/preprocessed/eeg/
tar -xzf ./analysis/eeg/data/preprocessed/eeg/preprocessed4.zip -C ./analysis/eeg/data/preprocessed/eeg/
tar -xzf ./analysis/eeg/data/preprocessed/eeg/preprocessed5.zip -C ./analysis/eeg/data/preprocessed/eeg/
tar -xzf ./analysis/eeg/data/processed/eeg/processed1.zip -C ./analysis/eeg/data/processed/eeg/
tar -xzf ./analysis/eeg/data/processed/eeg/processed2.zip -C ./analysis/eeg/data/processed/eeg/
tar -xzf ./analysis/eeg/data/results/results0.zip -C ./analysis/eeg/data/results/
tar -xzf ./analysis/spaces/spaces.zip -C ./analysis/spaces/

🗂️ Project structure

Your repository should now look roughly like this:

/root
│── /experiments/           # Source code for experiments
|──── /validation/          # Validation experiment
|────── /data/
|────── /resources/ 
|──── /online/              # Online experiment
|────── /data/
|────── /resources/ 
|──── /eeg/                 # EEG experiment
|────── /models/
|────── /resources/
|────── /rtfe/
|── /analysis/              # Source code for analysis
|──── /validation/          # Validation experiment
|────── /data/
|──── /online/              # Online experiment
|────── /data/
|──── /eeg/                 # EEG experiment
|────── /data/
|────── /rsa/
|──── /spaces/                      # Common packages and embeddings
|────── /pubplot/                   # Plotting utilities
|────── /embeddings/                # Embedding utilities
|────── /dumps/                     # Descriptive files for stimuli
|────── /glove-german/              # GloVe embeddings
|────── /text-embedding-ada-002/    # GPT3 embeddings
|────── /llama-7b/                  # LLama7b embeddings
|────── /bert-base-german-cased/    # BERT embeddings
│── README.md               # This file

⚙️ Usage

Here, we will walk through the basic steps of replicating all analyses reported in the paper and supplement. Not all steps that we report here will be necessary, if all you want to do is replicate the figures from our paper, as all processed data are included in OSF. This is because EEG analysis in particular is computationally expensive. For reference, we used two NVIDIA RTX 5000 Ada GPUs and the full analysis took about three weeks to run. Therefore, you may skip all optional steps if you would like to reproduce figures from processed results.

💡 Hint: Performance

Almost all computationally expensive modelling is done in the python scripts. For all of these, you may specify three key settings in your calls:

n_workers=n     How many jobs are we running in parallel?
backend=str     Which backend should be used (numpy/torch)?
device=str      If backend is torch, which device should be used (mps/cuda/cpu)?

These will dramatically influence computation times. Per default, all scripts will try to use torch with CUDA acceleration. Note that the default behaviour of these scripts is to distribute all available CUDA devices across your n_workers. If you would like only a specific device to be used, please specify which device should be used, e.g.:

python my_script.py n_workers=1 device=torch backend=cuda:0

Note that, if you are using GPUs, it is advisable to use as many workers as you have GPUs available (but no more---this may slow down computations considerably).

💡 Hint: Validate your installation

If you would like to fully replicate key results from our paper---specifically, refitting our models rather than using the provided solutions---you should now ensure that everything is running smoothly. To do so, please navigate to ./analysis/eeg/ in your terminal. Now, run:

conda activate sempriors
python
>>> import rsa
>>> rsa.math.unit_tests_2d()
>>> rsa.math.unit_tests_3d()
>>> rsa.analysis.estimators.unit_tests_TimeDelayed_trf()
>>> rsa.analysis.estimators.unit_tests_TimeDelayed_rec()
>>> rsa.analysis.estimators.unit_tests_Encoder()
>>> rsa.analysis.estimators.unit_tests_Temporal()
>>> rsa.analysis.estimators.unit_tests_SparseEncoder()
>>> rsa.analysis.estimators.unit_tests_Decoder()
>>> rsa.analysis.estimators.unit_tests_B2B()

For each of these, you should receive True as an output if tests are passed. Please repeat these tests for the torch backend like so, where you may optionally test a different device:

conda activate sempriors
python
>>> import rsa
>>> rsa.math.torch.unit_tests_2d(device = 'cpu')
>>> rsa.math.torch.unit_tests_3d(device = 'cpu')
>>> rsa.analysis.estimators.torch.unit_tests_TimeDelayed_trf(device = 'cpu')
>>> rsa.analysis.estimators.torch.unit_tests_TimeDelayed_rec(device = 'cpu')
>>> rsa.analysis.estimators.torch.unit_tests_Encoder(device = 'cpu')
>>> rsa.analysis.estimators.torch.unit_tests_Temporal(device = 'cpu')
>>> rsa.analysis.estimators.torch.unit_tests_SparseEncoder(device = 'cpu')
>>> rsa.analysis.estimators.torch.unit_tests_Decoder(device = 'cpu')
>>> rsa.analysis.estimators.torch.unit_tests_B2B(device = 'cpu')

❕ Optional: Validation experiment

Navigate to ./analysis/validation/. Open ./inference_validation.R, set the appropriate working directory in L17 and run the script. Results can be found in ./analysis/validation/results/items.csv.

❕ Optional: Online experiment

Navigate to ./analysis/online/. In your terminal, run:

python run_fe.py

Next, open a new jupyter notebook like so:

jupyter notebook

and open data.ipynb and follow the cells to aggregate all the data. Once you have run all cells, you may close the notebook and kernel. In RStudio, open./run_glmmm.R and adjust your working directory in L24. Run the code. Results will be available from ./analysis/online/results/glmm/.

❗️ Required: EEG experiment

This is where all analyses will come together and where figures will be created, hence I have marked this section as required. However, as mentioned above, many of these subsections will be computationally expensive. Individual sections that must be run will be marked again. Please make sure to navigate to ./analysis/eeg/.

💡 Hint: File names

For EEG analysis, files come in essentially four variants. These are:

/analysis/eeg/
|── audio_*.py          # Files associated with preprocessing audio data
|── subject_*.py        # Subject-level analysis scripts
|── group_*.py          # Group-level aggregation or analysis
|── inference_*         # Group-level statistical inference

Whenever you are interested about available parameters of or the inner workings of some kind of analysis, you will generally want to refer to the corresponding subject_* .py file that will contain descriptions of all parameters, usage notes, and the full code and documentation. Note that almost all group_* files are convenience scripts that will call individual subject_* scripts for all subjects.

❕ Optional: Computing gammatone spectrograms

Gammatone spectrograms for all stimuli can be extracted by running:

python audio_preprocess.py fs=200 fsn=200 method=spectrogram

❗️ Required: Computing semantic priors

During the experiment, we computed real-time estimates of semantic priors that participants had learned. Because these are computed deterministically and they require a relatively decent amount of space, these are not included in OSF. Please run:

python subject_rtfe.py
python subject_rtfe.py -generalised
python subject_rtfe.py -unspecific
python subject_rtfe.py -generalised -unspecific

These estimates will be used in all further scripts.

❕ Optional: Preprocessing EEG data

If you would like to re-run the preprocessing pipeline, you can run the following code. Note, however, that this is not a deterministic process (and some manual decisions must be taken). Consequently, you may opt to skip ahead. Otherwise, please run:

python subject_preprocess_rsa.py id=0002
python subject_preprocess_rerp.py id=0002

Both scripts will require user input for trial flagging, ICA, and final trial selection. Note that you will have to run these steps for all subjects (0002-0037). Note also that this will overwrite our preprocessed data.

❕ Optional: Behavioural analysis (task one)

Aggregate behavioural data by running:

python group_beh_mt1.py

Next, open inference_beh_mt1.R, adjust the working directory in L23 and run the script. Results are written to /analysis/eeg/data/processed/beh/mt1/. These will later be used during figure creation.

❗️ Required: Figure 1

By now, we may replicate figure one by opening a jupyter notebook:

jupyter notebook

and selecting ./fig1_design.ipynb. Run all cells consecutively. Figures will be saved to /analysis/eeg/figures/ as png, svg, and pdf files.

❕ Optional: Stimulus reconstruction

If you would like to run stimulus reconstruction over morphs, please run:

python group_rsa_rec.py n_workers=2 backend=torch device=cuda

Make sure to adjust the parameters as per performance hints. Note that you can also skip only this reconstruction step and move on from here: Once stimulus reconstruction has been completed (or skipped), you may run the inference script like so:

python inference_rsa_rec.py

Running this will compute all relevant statistics and tests. These will be available from /analysis/eeg/data/results/reconstruction.pkl.gz. Additionally, the script will generate an MNE report summarising key results that is available from /analysis/eeg/data/results/reports/reconstruction.html.

❕ Optional: Similarity encoding

If you would like to run similarity encoders from scratch, please run the following script, adjusting parameters as necessary:

python group_rsa_enc.py n_workers=2 backend=torch device=cuda b_acoustic=1 b_semantic=0
python group_rsa_enc.py n_workers=2 backend=torch device=cuda b_acoustic=0 b_semantic=1
python group_rsa_enc.py n_workers=2 backend=torch device=cuda b_acoustic=1 b_semantic=1
python inference_rsa_enc.py

Results from encoders will be available from /analysis/eeg/data/results/encoding_b0-m0-c0-k5.pkl.gz and, for the report, /analysis/eeg/data/results/reports/encoding_b0-m0-c0-k5.html.

Next, you may rerun encoders while systematically varying the number of top-k predictions considered like so:

python group_rsa_enc_topk.py n_workers=2 backend=torch device=cuda b_acoustic=1 b_semantic=0 n_mod=3
python inference_rsa_enc_topk.py

Results from top-k encoding will be available from /analysis/eeg/data/results/encoding_topk_b0-m0-c0.pkl.gzand /analysis/eeg/data/results/reports/encoding_topk_b0-m0-c0.html. In your report, check which number of k produced the best models. We know, of course, that this is going to be k=19, so let us now refit our original similarity encoders to make sure our results remain robust:

python group_rsa_enc.py n_workers=2 backend=torch device=cuda n_topk=19 b_acoustic=1 b_semantic=0
python group_rsa_enc.py n_workers=2 backend=torch device=cuda n_topk=19 b_acoustic=0 b_semantic=1
python group_rsa_enc.py n_workers=2 backend=torch device=cuda n_topk=19 b_acoustic=1 b_semantic=1
python inference_rsa_enc.py n_topk=19

Check your results in /analysis/eeg/data/results/encoding_b0-m0-c0-k19.pkl.gz and /analysis/eeg/data/results/reports/encoding_b0-m0-c0-k19.html.

❗️ Required: Figure 2

We now have all results available to reproduce figure two from the paper. If necessary, open your jupyter notebook

jupyter notebook

and select ./fig2_predictions.ipynb. Run all cells consecutively and find the resulting figure in /analysis/eeg/figures/ as png, svg, and pdf.

❕ Optional: Single-trial EEG encoding (task one)

First, please extract the full LLM activations---because these are quite large, it is genuinely faster to quickly extract them yourself rather than having to download them. Note that we do provide subspace projections, so you may also choose to skip this. To do this, please run

python audio_w2v2.py device=cuda folder=narrative
python audio_w2v2.py device=cuda folder=morphed
python audio_w2v2.py device=cuda folder=clear
python audio_w2v2.py device=cuda folder=vocoded

Next, perform the subspace projection like so:

python audio_w2v2_pca.py n_features=5
python audio_w2v2_pca.py n_features=10

Finally, run the back-to-back decoders to evaluate all layers:

python audio_w2v2_selection.py n_workers=2 backend=torch device=cuda n_features=5
python audio_w2v2_selection.py n_workers=2 backend=torch device=cuda n_features=10

Once this is completed, we may run the single-trial encoding models like so:

python group_rerp_mt1.py n_workers=2 backend=torch device=cuda s_bsl=llm n_features=5 s_mod=inv
python group_rerp_mt1.py n_workers=2 backend=torch device=cuda s_bsl=llm n_features=5 s_mod=spc
python inference_rerp_mt1.py s_bsl=llm n_features=5 s_mod=inv
python inference_rerp_mt1.py s_bsl=llm n_features=5 s_mod=spc

You may check results in /analysis/eeg/data/results/rerp-mt1-k5-z0-s0-b0-inv-llm5.pkl.gz, /analysis/eeg/data/results/rerp-mt1-k5-z0-s0-b0-spc-llm5.pkl.gz and, of course, reports in /analysis/eeg/data/results/reports/rerp-mt1-k5-z0-s0-b0-inv-llm5.html and /analysis/eeg/data/results/reports/rerp-mt1-k5-z0-s0-b0-spc-llm5.html. From these reports, find the best models and coefficients and run:

python group_rerp_mt1_knockout.py n_workers=2 backend=torch device=cuda s_bsl=llm n_features=5 s_mod=inv n_mod=2 a_coefs=all
python group_rerp_mt1_knockout.py n_workers=2 backend=torch device=cuda s_bsl=llm n_features=5 s_mod=spc n_mod=2 a_coefs=all
python inference_rerp_mt1_knockout.py s_bsl=llm n_features=5 s_mod=inv n_mod=2 a_coefs=all
python inference_rerp_mt1_knockout.py s_bsl=llm n_features=5 s_mod=spc n_mod=2 a_coefs=all

Results from knock-outs are available from /analysis/eeg/data/results/rerp-mt1-ko-n2-call-k5-z0-s0-b0-spc-llm5.pkl.gz, /analysis/eeg/data/results/rerp-mt1-ko-n2-call-k5-z0-s0-b0-inv-llm5.pkl.gz, and reports are available from /analysis/eeg/data/results/reports/rerp-mt1-ko-n2-call-k5-z0-s0-b0-inv-llm5.html and /analysis/eeg/data/results/reports/rerp-mt1-ko-n2-call-k5-z0-s0-b0-spc-llm5.html.

To replicate the robustness of these findings in the higher subspace projection and with respect to target words, please run:

python group_rerp_mt1.py n_workers=2 backend=torch device=cuda s_bsl=llm n_features=10 s_mod=inv
python group_rerp_mt1.py n_workers=2 backend=torch device=cuda s_bsl=llm n_features=10 s_mod=spc
python inference_rerp_mt1.py s_bsl=llm n_features=10 s_mod=inv
python inference_rerp_mt1.py s_bsl=llm n_features=10 s_mod=spc

python group_rerp_mt1_knockout.py n_workers=2 backend=torch device=cuda s_bsl=llm n_features=10 s_mod=inv n_mod=2 a_coefs=all
python group_rerp_mt1_knockout.py n_workers=2 backend=torch device=cuda s_bsl=llm n_features=10 s_mod=spc n_mod=2 a_coefs=all
python inference_rerp_mt1_knockout.py s_bsl=llm n_features=10 s_mod=inv n_mod=2 a_coefs=all
python inference_rerp_mt1_knockout.py s_bsl=llm n_features=10 s_mod=spc n_mod=2 a_coefs=all

and:

python group_rerp_mt1.py n_workers=2 backend=torch device=cuda s_bsl=tar s_mod=inv
python group_rerp_mt1.py n_workers=2 backend=torch device=cuda s_bsl=tar s_mod=spc
python inference_rerp_mt1.py s_bsl=tar s_mod=inv
python inference_rerp_mt1.py s_bsl=tar s_mod=spc

python group_rerp_mt1_knockout.py n_workers=2 backend=torch device=cuda s_bsl=tar s_mod=inv n_mod=2 a_coefs=all
python group_rerp_mt1_knockout.py n_workers=2 backend=torch device=cuda s_bsl=tar s_mod=spc n_mod=2 a_coefs=all
python inference_rerp_mt1_knockout.py s_bsl=tar s_mod=inv n_mod=2 a_coefs=all
python inference_rerp_mt1_knockout.py s_bsl=tar s_mod=spc n_mod=2 a_coefs=all

❗️ Required: Figure 3

We can now reproduce figure three from the paper. Again, open your jupyter notebook:

jupyter notebook

and select ./fig3_surprisal.ipynb. Run all cells consecutively and find the figures in /analysis/eeg/figures/ as png, svg, and pdf.

❕ Optional: Modeling response times by congruency (task two)

First, aggregate all behavioural data from task two:

python group_beh_mt2.py

Next, open /analysis/eeg/inference_beh_mt2.R, adjust the working directory in L23 and run the script. Outputs are available from /analysis/eeg/data/processed/beh/mt2/ and will be used for plotting.

❕ Optional: Single-trial EEG encoding (task two)

To replicate single-trial EEG encoding in task two, please run:

python group_rerp_mt2.py n_workers=2 backend=torch device=cuda b_con=1 b_inc=0 s_mod=inv
python group_rerp_mt2.py n_workers=2 backend=torch device=cuda b_con=0 b_inc=1 s_mod=inv
python group_rerp_mt2.py n_workers=2 backend=torch device=cuda b_con=1 b_inc=0 s_mod=spc
python group_rerp_mt2.py n_workers=2 backend=torch device=cuda b_con=0 b_inc=1 s_mod=spc

python inference_rerp_mt2.py b_con=1 b_inc=0 s_mod=inv
python inference_rerp_mt2.py b_con=0 b_inc=1 s_mod=inv
python inference_rerp_mt2.py b_con=1 b_inc=0 s_mod=spc
python inference_rerp_mt2.py b_con=0 b_inc=1 s_mod=spc

Outputs will be available from /analysis/eeg/data/results/rerp-mt2-c0-i1-k5-z0-s0-b0-inv.pkl.gz, /analysis/eeg/data/results/rerp-mt2-c0-i1-k5-z0-s0-b0-spc.pkl.gz, /analysis/eeg/data/results/rerp-mt2-c1-i0-k5-z0-s0-b0-inv.pkl.gz/, /analysis/eeg/data/results/rerp-mt2-c1-i0-k5-z0-s0-b0-spc.pkl.gz, and, of course, the corresponding reports in /analysis/eeg/data/results/reports/.

Again, please check the reports to identify the best performing models each. We know that these will be speaker-invariant and speaker-specific semantic surprisal for incongruent and congruent trials, respectively. Hence, we can run the following:

python group_rerp_mt2_knockout.py n_workers=2 backend=torch device=cuda b_con=0 b_inc=1 s_mod=inv n_mod=2 a_coefs=all
python group_rerp_mt2_knockout.py n_workers=2 backend=torch device=cuda b_con=1 b_inc=0 s_mod=spc n_mod=2 a_coefs=all

python inference_rerp_mt2_knockout.py b_con=0 b_inc=1 s_mod=inv n_mod=2 a_coefs=all
python inference_rerp_mt2_knockout.py b_con=1 b_inc=0 s_mod=spc n_mod=2 a_coefs=all

This will generate our final output data in /analysis/eeg/data/results/rerp-mt2-ko-m2-c0-i1-k5-z0-s0-b0-inv.pkl.gz, /analysis/eeg/data/results/rerp-mt2-ko-m2-c1-i0-k5-z0-s0-b0-spc.pkl.gz, and in /analysis/eeg/data/results/reports/.

❗️ Required: Figure 4

We have now replicated all analysis steps from the paper. To recreate figure four, let's open our jupyter notebook again:

jupyter notebook

Select ./fig4_limits.ipynb and run all cells. Figures will be exported to /analysis/eeg/figures/ as png, svg, and pdf.

❗️ Required: Supplementary Figures and Tables

At this point, we have also run all steps of control analysis already. Consequently, we can already generate supplementary results. We will not go through each individual file here, but they are all available in notebooks following suppfig_*.ipynb. Running all cells within these notebooks will output figures to /analysis/eeg/figures/ and will output a neat LaTeX table with key results that need to be reported.

🪣 Clean-up

Once you've replicated our results and are ready to move on, don't forget to clean-up your hard-drive. You can simply remove the root directory manually and run:

conda deactivate sempriors
conda remove -n sempriors --all