diff --git a/data-wiki/README.md b/data-wiki/README.md
index 9140ddf..257f42d 100644
--- a/data-wiki/README.md
+++ b/data-wiki/README.md
@@ -66,8 +66,42 @@ parallel --verbose -j 8 ./convert-doc.sh '<' {} '>' {.}.md \
   < wiki/docs.txt \
   2>&1 | tee convert.log
 ```
+This may take a few days.
 
 There are some downsides to using Pandoc here, since it does not handle Wikipedia template
 references, and instead seems to remove them in the output. This leads to a few sentences
 missing words in the middle. This is a relatively rare occasion, so it should not be much
 of a problem.
+
+Also, the conversion crashes sometimes:
+```
+/home/leod/src/hncynic/data-wiki/convert-doc.sh < docs/85/Munich%E2%80%93Augsburg_railway.txt > docs/85/Munich%E2%80%93Augsburg_railway.md
+Traceback (most recent call last):
+  File "/home/leod/src/hncynic/data-wiki/filter_markdown.py", line 114, in <module>
+    main()
+  File "/home/leod/src/hncynic/data-wiki/filter_markdown.py", line 98, in main
+    return run_filter(action, prepare=prepare, doc=doc)
+  File "/home/leod/.local/lib/python3.6/site-packages/panflute/io.py", line 260, in run_filter
+    return run_filters([action], *args, **kwargs)
+...
+  File "/home/leod/.local/lib/python3.6/site-packages/panflute/elements.py", line 1061, in __init__
+    self.header = header
+  File "/home/leod/.local/lib/python3.6/site-packages/panflute/elements.py", line 1097, in header
+    raise IndexError(msg)
+IndexError: table header has an incorrect number of cols: 6 rows but expected 8
+pandoc: Error running filter /home/leod/src/hncynic/data-wiki/filter_markdown.py
+Filter returned error status 1
+```
+## Convert to TSV
+We use each section of an article as an individual training example.
+
+```
+find docs -name '*.md' > docs.md.txt
+parallel --verbose -j 8 \
+  ~/src/hncynic/data-wiki/clean_text.sh \
+    '<' {} \
+    '|' ~/src/hncynic/data-wiki/md_to_tsv.py {} \
+    '>' {.}.tsv \
+  < docs.md.txt \
+  > convert.tsv.log 2>&1
+```
diff --git a/data-wiki/md_to_tsv.py b/data-wiki/md_to_tsv.py
index 4d451e8..4f085a6 100755
--- a/data-wiki/md_to_tsv.py
+++ b/data-wiki/md_to_tsv.py
@@ -41,7 +41,7 @@ def write_tsv(self, out, header_path=None):
       out.write('\t')
       for i in range(len(self.text)-1):
         out.write(self.text[i].strip())
-        out.write(' <P> ')
+        out.write(' <NL> <NL> ')
       out.write(self.text[-1])
       out.write('\n')
 
@@ -98,7 +98,6 @@ def parse_md_tree(stream, title):
 
   title = urllib.parse.unquote(basename.replace('_', ' '))
 
-  with open(filename) as f:
-    tree = parse_md_tree(f, title)
-    #tree.write(sys.stdout)
-    tree.write_tsv(sys.stdout)
+  tree = parse_md_tree(sys.stdin, title)
+  #tree.write(sys.stdout)
+  tree.write_tsv(sys.stdout)
diff --git a/data/moses_tokenizer_protected.txt b/data/moses_tokenizer_protected.txt
index 599dbaa..b67112e 100644
--- a/data/moses_tokenizer_protected.txt
+++ b/data/moses_tokenizer_protected.txt
@@ -1,4 +1,5 @@
 <NL>
+<P>
 http://\S+
 https://\S+
 ftp://\S+
diff --git a/train/README.md b/train/README.md
index 68404f9..869394c 100644
--- a/train/README.md
+++ b/train/README.md
@@ -36,9 +36,16 @@ the power plug.
 ```
 onmt-main train --config opennmt_config.yml --model_type Transformer --num_gpus 1
 ```
+I trained the model for about 40K steps with `opennmt_config.yml`. I noticed that the loss wasn't
+improving much after that, so I got worried and increased the batch size (known to help with
+training Transformer models) by performing gradient accumulation as in `opennmt_config_larger_batch.yml`.
+As can be seen in the plot below, this seems to have helped.
 
 ![training loss](train.svg)
 
+Unfortunately, I don't have a plot for the dev loss, since I forgot to turn on dev evaluation.
+What a bummer.
+
 ### Evaluate
 
 ### Export