diff --git a/.gitignore b/.gitignore
index 7f73e11..19976c2 100644
--- a/.gitignore
+++ b/.gitignore
@@ -4,6 +4,9 @@
 networks/
 checkpoints/
 
+# Image used for quick smoke tests
+82148729_p0.jpg
+
 # Bunch of files that have no place in version control
 2020_0000_0599/encoded_tags_test.npy
 2020_0000_0599/stats.txt
diff --git a/README.md b/README.md
index 88d6bd9..2b0b2e0 100644
--- a/README.md
+++ b/README.md
@@ -6,3 +6,43 @@ Repo for my Tensorflow/Keras CV experiments. Mostly revolving around the Danboor
 Framework: TF/Keras 2.7
 
 Training SQLite DB built using fire-egg's tools: https://github.com/fire-eggs/Danbooru2019
+
+Currently training on Danbooru2021, 512px SFW subset (sans the rating:q images that had been included in the 2022-01-21 release of the dataset)
+
+## Reference:
+Anonymous, The Danbooru Community, & Gwern Branwen; “Danbooru2021: A Large-Scale Crowdsourced and Tagged Anime Illustration Dataset”, 2022-01-21. Web. Accessed 2022-01-28 https://www.gwern.net/Danbooru2021
+
+----
+
+## Journal
+06/02/2022: great news crew! TRC allowed me to use a bunch of TPUs!
+
+To make better use of this amount of compute I had to overhaul a number of components, so a bunch of things are likely to have fallen to bitrot in the process.
+I can only guarantee NFNet can work pretty much as before with the right arguments.
+NFResNet changes *should* have left it retrocompatible with the previous version.
+ResNet has been streamlined to be mostly in line with the Bag-of-tricks paper ([arXiv:1812.01187](https://arxiv.org/abs/1812.01187)) with the exception of the stem. It is not compatible with the previous version of the code.
+
+The training labels have been included in the 2021_0000_0899 folder for convenience.
+The list of files used for training is going to be uploaded as a GitHub Release.
+
+Now for some numbers:
+compared to my previous best run, the one that resulted in [NFNetL1V1-100-0.57141](https://github.com/SmilingWolf/SW-CV-ModelZoo/releases/tag/NFNetL1V1-100-0.57141):
+- I'm using 1.86x the amount of images: 2.8M vs 1.5M
+- I'm training bigger models: 61M vs 45M params
+- ... in less time: 232 vs 700 hours of processor time
+- don't get me started on actual wall clock time
+- with a few amenities thrown in: ECA for channel attention, SiLU activation
+
+And it's all thanks to the folks at TRC, so shout out to them!
+
+I currently have a few runs in progress across a couple of dimensions:
+- effect of model size with NFNet L0/L1/L2, with SiLU and ECA for all three of them
+- effect of activation function with NFNet L0, with SiLU/HSwish/ReLU, no ECA
+
+Once the experiments are over, the plan is to select the network definitions that lay on the Pareto curve between throughput and F1 score and release the trained weights.
+
+One last thing.
+I'd like to call your attention to the tools/cleanlab_stuff.py script.
+It reads two files: one with the binarized labels from the database, the other with the predicted probabilities.
+It then uses the [cleanlab](https://github.com/cleanlab/cleanlab) package to estimate whether if an image in a set could be missing a given label. At the end it stores its conclusions in a json file.
+This file could, potentially, be used in some tool to assist human intervention to add the missing tags.
diff --git a/tools/cleanlab_stuff.py b/tools/cleanlab_stuff.py
index 09f1b08..90e56ec 100644
--- a/tools/cleanlab_stuff.py
+++ b/tools/cleanlab_stuff.py
@@ -16,7 +16,7 @@
         samples = [x.rstrip() for x in f.readlines()]
 
     full_labels = np.load("2021_0000_0899/encoded_tags_test.npy")
-    full_psx = np.load("tags_probs_NFNetL0V1_01_29_2022_08h29m32s.npy")
+    full_psx = np.load("tags_probs_NFNetL1V1_01_29_2022_08h20m44s.npy")
 
     tags_actions = {}
 
@@ -32,7 +32,10 @@
         psx = np.stack([recip, psx], axis=1)
 
         ordered_label_errors = get_noise_indices(
-            s=train_labels_with_errors, psx=psx, sorted_index_method="normalized_margin"
+            s=train_labels_with_errors,
+            psx=psx,
+            sorted_index_method="normalized_margin",
+            n_jobs=1,
         )
 
         # On Danbooru a tag is more likely to be missing than to be wrong,