From 5d2a8b31be17a2992308d58b674274375a3f360c Mon Sep 17 00:00:00 2001 From: oobabooga <112222186+oobabooga@users.noreply.github.com> Date: Thu, 29 Jun 2023 14:33:47 -0300 Subject: [PATCH] Improve Parameters tab UI --- server.py | 58 ++++++++++++++++++++++++++++++++++++++++++------------- 1 file changed, 45 insertions(+), 13 deletions(-) diff --git a/server.py b/server.py index 9fdbfd92..80198e3a 100644 --- a/server.py +++ b/server.py @@ -327,24 +327,56 @@ def create_settings_menus(default_preset): gr.Markdown('Main parameters') with gr.Row(): with gr.Column(): - shared.gradio['temperature'] = gr.Slider(0.01, 1.99, value=generate_params['temperature'], step=0.01, label='temperature', info='Primary factor to control randomness of outputs. 0 = deterministic (only the most likely token is used). Higher value = more randomness.') - shared.gradio['top_p'] = gr.Slider(0.0, 1.0, value=generate_params['top_p'], step=0.01, label='top_p', info='If not set to 1, select tokens with probabilities adding up to less than this number. Higher value = higher range of possible random results.') - shared.gradio['top_k'] = gr.Slider(0, 200, value=generate_params['top_k'], step=1, label='top_k', info='Similar to top_p, but select instead only the top_k most likely tokens. Higher value = higher range of possible random results.') - shared.gradio['typical_p'] = gr.Slider(0.0, 1.0, value=generate_params['typical_p'], step=0.01, label='typical_p', info='If not set to 1, select only tokens that are at least this much more likely to appear than random tokens, given the prior text.') - shared.gradio['epsilon_cutoff'] = gr.Slider(0, 9, value=generate_params['epsilon_cutoff'], step=0.01, label='epsilon_cutoff', info='In units of 1e-4; a reasonable value is 3. This sets a probability floor below which tokens are excluded from being sampled. Should be used with top_p, top_k, and eta_cutoff set to 0.') - shared.gradio['eta_cutoff'] = gr.Slider(0, 20, value=generate_params['eta_cutoff'], step=0.01, label='eta_cutoff', info='In units of 1e-4; a reasonable value is 3. Should be used with top_p, top_k, and epsilon_cutoff set to 0.') + shared.gradio['temperature'] = gr.Slider(0.01, 1.99, value=generate_params['temperature'], step=0.01, label='temperature') + shared.gradio['top_p'] = gr.Slider(0.0, 1.0, value=generate_params['top_p'], step=0.01, label='top_p') + shared.gradio['top_k'] = gr.Slider(0, 200, value=generate_params['top_k'], step=1, label='top_k') + shared.gradio['typical_p'] = gr.Slider(0.0, 1.0, value=generate_params['typical_p'], step=0.01, label='typical_p') + shared.gradio['epsilon_cutoff'] = gr.Slider(0, 9, value=generate_params['epsilon_cutoff'], step=0.01, label='epsilon_cutoff') + shared.gradio['eta_cutoff'] = gr.Slider(0, 20, value=generate_params['eta_cutoff'], step=0.01, label='eta_cutoff') with gr.Column(): - shared.gradio['repetition_penalty'] = gr.Slider(1.0, 1.5, value=generate_params['repetition_penalty'], step=0.01, label='repetition_penalty', info='Exponential penalty factor for repeating prior tokens. 1 means no penalty, higher value = less repetition, lower value = more repetition.') - shared.gradio['repetition_penalty_range'] = gr.Slider(0, 4096, step=64, value=generate_params['repetition_penalty_range'], label='repetition_penalty_range', info='The number of most recent tokens to consider for repetition penalty. 0 makes all tokens be used.') - shared.gradio['encoder_repetition_penalty'] = gr.Slider(0.8, 1.5, value=generate_params['encoder_repetition_penalty'], step=0.01, label='encoder_repetition_penalty', info='Also known as the "Hallucinations filter". Used to penalize tokens that are *not* in the prior text. Higher value = more likely to stay in context, lower value = more likely to diverge.') - shared.gradio['no_repeat_ngram_size'] = gr.Slider(0, 20, step=1, value=generate_params['no_repeat_ngram_size'], label='no_repeat_ngram_size', info='If not set to 0, specifies the length of token sets that are completely blocked from repeating at all. Higher values = blocks larger phrases, lower values = blocks words or letters from repeating. Only 0 or high values are a good idea in most cases.') - shared.gradio['min_length'] = gr.Slider(0, 2000, step=1, value=generate_params['min_length'], label='min_length', info='Minimum generation length in tokens.') + shared.gradio['repetition_penalty'] = gr.Slider(1.0, 1.5, value=generate_params['repetition_penalty'], step=0.01, label='repetition_penalty') + shared.gradio['repetition_penalty_range'] = gr.Slider(0, 4096, step=64, value=generate_params['repetition_penalty_range'], label='repetition_penalty_range') + shared.gradio['encoder_repetition_penalty'] = gr.Slider(0.8, 1.5, value=generate_params['encoder_repetition_penalty'], step=0.01, label='encoder_repetition_penalty') + shared.gradio['no_repeat_ngram_size'] = gr.Slider(0, 20, step=1, value=generate_params['no_repeat_ngram_size'], label='no_repeat_ngram_size') + shared.gradio['min_length'] = gr.Slider(0, 2000, step=1, value=generate_params['min_length'], label='min_length') shared.gradio['tfs'] = gr.Slider(0.0, 1.0, value=generate_params['tfs'], step=0.01, label='tfs') shared.gradio['top_a'] = gr.Slider(0.0, 1.0, value=generate_params['top_a'], step=0.01, label='top_a') shared.gradio['do_sample'] = gr.Checkbox(value=generate_params['do_sample'], label='do_sample') - gr.Markdown('[Click here for more information.](https://github.com/oobabooga/text-generation-webui/blob/main/docs/Generation-parameters.md)') + with gr.Accordion("Learn more", open=False): + gr.Markdown(""" + + Not all parameters are used by all loaders. See [this page](https://github.com/oobabooga/text-generation-webui/blob/main/docs/Generation-parameters.md) for details. + + For a technical description of the parameters, the [transformers documentation](https://huggingface.co/docs/transformers/main_classes/text_generation#transformers.GenerationConfig) is a good reference. + + ### Temperature + Primary factor to control randomness of outputs. 0 = deterministic (only the most likely token is used). Higher value = more randomness. + ### top_p + If not set to 1, select tokens with probabilities adding up to less than this number. Higher value = higher range of possible random results. + ### top_k + Similar to top_p, but select instead only the top_k most likely tokens. Higher value = higher range of possible random results. + ### typical_p + If not set to 1, select only tokens that are at least this much more likely to appear than random tokens, given the prior text. + ### epsilon_cutoff + In units of 1e-4; a reasonable value is 3. This sets a probability floor below which tokens are excluded from being sampled. Should be used with top_p, top_k, and eta_cutoff set to 0. + ### eta_cutoff + In units of 1e-4; a reasonable value is 3. Should be used with top_p, top_k, and epsilon_cutoff set to 0. + ### repetition_penalty + Exponential penalty factor for repeating prior tokens. 1 means no penalty, higher value = less repetition, lower value = more repetition. + ### repetition_penalty_range + The number of most recent tokens to consider for repetition penalty. 0 makes all tokens be used. + ### encoder_repetition_penalty + Also known as the "Hallucinations filter". Used to penalize tokens that are *not* in the prior text. Higher value = more likely to stay in context, lower value = more likely to diverge. + ### no_repeat_ngram_size + If not set to 0, specifies the length of token sets that are completely blocked from repeating at all. Higher values = blocks larger phrases, lower values = blocks words or letters from repeating. Only 0 or high values are a good idea in most cases. + ### min_length + Minimum generation length in tokens. + ### penalty_alpha + Contrastive Search is enabled by setting this to greater than zero and unchecking "do_sample". It should be used with a low value of top_k, for instance, top_k = 4. + + """, elem_classes="markdown") with gr.Column(): create_chat_settings_menus() @@ -352,7 +384,7 @@ def create_settings_menus(default_preset): with gr.Row(): with gr.Column(): gr.Markdown('Contrastive search') - shared.gradio['penalty_alpha'] = gr.Slider(0, 5, value=generate_params['penalty_alpha'], label='penalty_alpha', info='Contrastive Search is enabled by setting this to greater than zero and unchecking "do_sample". It should be used with a low value of top_k, for instance, top_k = 4.') + shared.gradio['penalty_alpha'] = gr.Slider(0, 5, value=generate_params['penalty_alpha'], label='penalty_alpha') gr.Markdown('Beam search') shared.gradio['num_beams'] = gr.Slider(1, 20, step=1, value=generate_params['num_beams'], label='num_beams')