[kosha] Substantially increase size and coverage

This commit uses data from the Upasargartha-candrika to create a large
number of prefixed tinantas and krdantas.

In addition, I've fixed a few minor bugs and added more documentation to
these crates.

Makefile

@@ -37,7 +37,9 @@ create_sandhi_rules:
 # Creates a koshas and write it to disk.
 create_kosha:
 	RUST_LOG=info cargo run --release --bin create_kosha -- \
-			 --input-dir data/raw/lex --output-dir data/build/vidyut-latest
+			 --input-dir data/raw/lex \
+			 --dhatupatha vidyut-prakriya/data/dhatupatha.tsv \
+			 --output-dir data/build/vidyut-latest
 
 # Trains a padaccheda model and saves important features to disk.
 # NOTE: when training, exclude the file paths used in `make eval`.

README.md

@@ -110,7 +110,9 @@ In Rust, components of this kind are called *crates*.
 
 ### [`vidyut-chandas`][vidyut-chandas]
 
-`vidyut-chandas` is an experimental classifier for Sanskrit meters.
+`vidyut-chandas` identifies the meter in some piece of Sanskrit text. This
+crate is experimental, and while it is useful for common and basic use cases,
+it is not a state-of-the-art solution.
 
 For details, see the [vidyut-chandas README][vidyut-chandas].
 

scripts/create_all_data.sh

@@ -22,6 +22,8 @@ else
     echo "Training data does not exist -- fetching."
     mkdir -p "data/raw/dcs"
     git clone --depth 1 https://github.com/OliverHellwig/sanskrit.git dcs-data
+    # Use a fixed commit to avoid breakages from later changes.
+    pushd dcs-data && git reset --hard 1bc281e && popd
     mv dcs-data/dcs/data/conllu data/raw/dcs/conllu
     rm -Rf dcs-data
 fi