Update medium-cpu.cfg

[Korretdesigns]

###################

General

###################

[general_params]

Device ordinal for CPU/GPU supports.

Setting this to -1 will leverage CPU, a positive will run the model on the associated CUDA device id.

device = -1

Seed for random number generators, fix seed to reproduce results.

By default there is no seed and each turn should be unique.

seed

Whether to enable debugging.

debug = False

###########################

Text generation

###########################

DialoGPT (see https://github.com/microsoft/DialoGPT) or any other text generator supported by transformers.pipeline.

[generation_pipeline_kwargs]

Keyword arguments passed to the text generation pipeline.

The model that will be used by the pipeline 'text-generation' to generate responses.

The model must be an instance of a pretrained model inheriting from PreTrainedModel or TFPreTrainedModel.

model = microsoft/DialoGPT-medium

The configuration that will be used by the pipeline to instantiate the model.

config

The tokenizer that will be used by the pipeline to encode data for the model.

tokenizer

The framework to use, either pt for PyTorch or tf for TensorFlow.

The specified framework must be installed.

framework

[generator_kwargs]

Keyword arguments passed to the text generator.

See https://huggingface.co/blog/how-to-generate

The maximal number of tokens to be returned, inclusive of punctuations etc.

It will automatically stop if the end-of-sequence token was found earlier.

max_length = 1000

The minimum length of the sequence to be generated.

min_length = 1

Whether or not to use sampling; use greedy decoding otherwise.

do_sample = True

Whether to stop the beam search when at least num_beams sentences are finished per batch or not.

early_stopping = False

Number of beams for beam search.

1 means no beam search.

num_beams = 1

Number of groups to divide num_beams into in order to ensure diversity among different groups of beams.

1 means no group beam search.

num_beam_groups = 1

Value to control diversity for group beam search.

The higher the penalty, the more diverse are the outputs.

diversity_penalty = 0.0

The value used to module the next token probabilities.

Must be strictly positive.

Lower temperature results in less random completions.

As the temperature approaches zero, the model will become deterministic and repetitive.

Higher temperature results in more random completions.

temperature = 1

The number of highest probability vocabulary tokens to keep for top-k-filtering.

1 means only 1 word is considered for each step (token), resulting in deterministic completions,

while 40 means 40 words are considered at each step.

0 (default) is a special setting meaning no restrictions.

40 generally is a good value.

top_k = 40

If set to float < 1, only the most probable tokens with probabilities

that add up to top_p or higher are kept for generation.

top_p = 0.9

The parameter for repetition penalty. 1.0 means no penalty.

repetition_penalty = 1

Exponential penalty to the length. 1.0 means no penalty.

Set to values < 1.0 in order to encourage the model to generate shorter sequences,

to a value > 1.0 in order to encourage the model to produce longer sequences.

length_penalty = 1

If set to int > 0, all ngrams of that size can only occur once.

no_repeat_ngram_size = 0

The id of the padding token.

pad_token_id

The id of the beginning-of-sequence token.

bos_token_id

The id of the end-of-sequence token.

eos_token_id

Comma separated list of token ids that are not allowed to be generated.

bad_words_ids

The number of independently computed returned sequences for each element in the batch.

You would need to use a response classifier or implement a function.

For example, you can choose the most dissimilar message, or the lengthiest one.

But keep in mind: the higher the number, the slower the generation.

num_return_sequences = 1

If an encoder-decoder model starts decoding with a different token than bos, the id of that token.

decoder_start_token_id

Whether or not the model should use the past last key/values attentions

(if applicable to the model) to speed up decoding.

use_cache = True

Whether or not to clean up the potential extra spaces in the text output.

clean_up_tokenization_spaces = True

############################

Response ranking

############################

DialogRPT (see https://github.com/golsun/DialogRPT)

Any ranker can be disabled by setting to empty.

NOTE: Ensure num_return_sequences is greater than 1.

[prior_ranker_weights]

The prior score is the weighted average of human_vs_rand and human_vs_machine predictions.

Weight of "How relevant the response is for the given context?"

human_vs_rand_weight

Weight of "How likely the response is human-written rather than machine-generated?"

human_vs_machine_weight

[cond_ranker_weights]

The cond score is the weighted average of updown, depth and width predictions.

Weight of "How likely the response gets the most upvotes?"

updown_weight

Weight of "How likely the response gets the most direct replies?"

depth_weight

Weight of "How likely the response gets the longest follow-up thread?"

width_weight

#########################

Communication

#########################

[chatbot_params]

Parameters of the chatbot itself.

The number of turns (turn = answer and response) the model should consider.

Set to 0 to focus on the last message. Set to -1 for unlimited context length.

max_turns_history = 2

Your Telegram token. See https://core.telegram.org/bots

telegram_token = 7174689615:AAH9BptKSkuNmvAnuesLUs9oDOyzHsEL3UE

Your GIPHY API token. See https://developers.giphy.com/docs/api/

giphy_token = SgNvjECWmHeOfYsnUWlAfaeWYkqwQu83

Probability of returning a GIF.

giphy_prob = 0.1

The maximal number of words the bot has to generate to also return a GIF.

giphy_max_words = 3

Value from 0-10 which makes results weirder as you go up the scale.

giphy_weirdness = 5

Whether to continue from the previous dialogue.

continue_after_restart = True

The filename for storing the pickle files.

data_filename = bot_data.pkl