pangoling 1.0.3
- Internal changes to comply with CRAN requirements.
- HF_HOME is used now to store the models rather than
TRANSFORMERS_CACHE
pangoling 1.0.2
- Internal changes: OMP THREAD LIMIT was set to 1.
pangoling 1.0.1
New Features
- Added
installed_py_pangoling() to check if required
Python dependencies (transformers and torch)
are installed.
Other changes
- Informative startup message if python dependencies not
installed.
- Documentation examples won’t run if python dependencies not
installed
- Articles are now pre-computed vignettes. See
pangoling 1.0.0
- changed the ownership of the repo to ropensci
- deprecated functions are now defunct and have been replaced with
their respective alternative functions
pangoling 0.0.0.9011
- Added
word_n argument in
causal_words_pred() to indicate word order of the
texts.
- Allows for models with larger vocabulary than tokenizer.
pangoling 0.0.0.9010
New Features:
- Added
checkpoint parameter to
causal_preload() and masked_preload() to allow
loading models from checkpoints.
- Introduced
causal_next_tokens_pred_tbl(), which
replaces causal_next_tokens_tbl() and provides improved
predictability calculations.
- Added
causal_words_pred(),
causal_targets_pred(), and
causal_tokens_pred_lst() to compute predictability for
words, phrases, or tokens, replacing causal_lp() and
causal_tokens_lp_tbl().
- Introduced
masked_tokens_pred_tbl(), replacing
masked_tokens_tbl(), for retrieving possible tokens and
their log probabilities.
- Introduced
masked_targets_pred(), replacing
masked_lp(), for calculating predictability based on left
and right context.
- Introduced
transformer_vocab() with an optional
decode parameter to return decoded tokenized words.
- New dataset
df_jaeger14: Self-paced
reading data on Chinese relative clauses.
- New dataset
df_sent: Example dataset
with two word-by-word sentences.
- New vignette: Added a worked-out example of a
causal model.
Enhancements:
- Added
sep argument in causal_words_pred()
to support languages without spaces between words (e.g., Chinese).
- New
log.p argument across multiple functions to specify
how predictability is calculated (e.g., log base e, log base 2
for bits, or raw probabilities).
- Improved tokenization utilities:
tokenize_lst() now
supports decoded outputs via the decode parameter.
- Updated
install_py_pangoling() to enhance Python
environment handling.
- Added
perplexity_calc() for computing perplexity from
probabilities.
Deprecations:
- Deprecated
causal_next_tokens_tbl(),
causal_lp(), causal_tokens_lp_tbl(), and
causal_lp_mats(). Use
causal_next_tokens_pred_tbl(),
causal_targets_pred(), causal_words_pred(),
and causal_pred_mats() instead.
- Deprecated
masked_tokens_tbl() and
masked_lp(). Use masked_tokens_pred_tbl() and
masked_targets_pred() instead.
pangoling 0.0.0.9009
- Deprecated
.by in favor of by.
pangoling 0.0.0.9008
- Fix a bug when
.by is unordered
pangoling 0.0.0.9007
set_cache_folder() function added.
- Message when the package loads.
- New troubleshooting vignette.
pangoling 0.0.0.9006
causal_lp get a l_contexts argument.
- Checkpoints work for causal models (not yet for masked models).
- Ropensci badge added.
pangoling 0.0.0.9005
- Strings with no tokens no longer throw errors.
- Requires correct version of R.
pangoling 0.0.0.9004
- Causal models accept batches.
pangoling 0.0.0.9003
- bug in causal_tokens_lp_tbl fixed
pangoling 0.0.0.9002
- minor function names to avoid conflict with other packages
pangoling 0.0.0.9001
- Tons of stuff. Fully functional package now.
pangoling 0.0.0.9000