@huggingface/tasks 0.2.1 → 0.3.0

package/src/tasks/text-generation/data.ts

@@ -12,12 +12,12 @@ const taskData: TaskDataCustom = {
 			id: "the_pile",
 		},
 		{
-			description: "A crowd-sourced instruction dataset to develop an AI assistant.",
-			id: "OpenAssistant/oasst1",
+			description: "Truly open-source, curated and cleaned dialogue dataset.",
+			id: "HuggingFaceH4/ultrachat_200k",
 		},
 		{
-			description: "A crowd-sourced instruction dataset created by Databricks employees.",
-			id: "databricks/databricks-dolly-15k",
+			description: "An instruction dataset with preference ratings on responses.",
+			id: "openbmb/UltraFeedback",
 		},
 	],
 	demo: {
@@ -59,66 +59,50 @@ const taskData: TaskDataCustom = {
 			id: "bigcode/starcoder",
 		},
 		{
-			description: "A model trained to follow instructions, uses Pythia-12b as base model.",
-			id: "databricks/dolly-v2-12b",
+			description: "A very powerful text generation model.",
+			id: "mistralai/Mixtral-8x7B-Instruct-v0.1",
 		},
 		{
-			description: "A model trained to follow instructions curated by community, uses Pythia-12b as base model.",
-			id: "OpenAssistant/oasst-sft-4-pythia-12b-epoch-3.5",
+			description: "Small yet powerful text generation model.",
+			id: "microsoft/phi-2",
 		},
 		{
-			description: "A large language model trained to generate text in English.",
-			id: "stabilityai/stablelm-tuned-alpha-7b",
+			description: "A very powerful model that can chat, do mathematical reasoning and write code.",
+			id: "openchat/openchat-3.5-0106",
 		},
 		{
-			description: "A model trained to follow instructions, based on mosaicml/mpt-7b.",
-			id: "mosaicml/mpt-7b-instruct",
+			description: "Very strong yet small assistant model.",
+			id: "HuggingFaceH4/zephyr-7b-beta",
 		},
 		{
-			description: "A large language model trained to generate text in English.",
-			id: "EleutherAI/pythia-12b",
-		},
-		{
-			description: "A large text-to-text model trained to follow instructions.",
-			id: "google/flan-ul2",
-		},
-		{
-			description: "A large and powerful text generation model.",
-			id: "tiiuae/falcon-40b",
-		},
-		{
-			description: "State-of-the-art open-source large language model.",
+			description: "Very strong open-source large language model.",
 			id: "meta-llama/Llama-2-70b-hf",
 		},
 	],
 	spaces: [
 		{
-			description: "A robust text generation model that can perform various tasks through natural language prompting.",
-			id: "huggingface/bloom_demo",
+			description: "A leaderboard to compare different open-source text generation models based on various benchmarks.",
+			id: "HuggingFaceH4/open_llm_leaderboard",
 		},
 		{
-			description: "An text generation based application that can write code for 80+ languages.",
-			id: "bigcode/bigcode-playground",
+			description: "An text generation based application based on a very powerful LLaMA2 model.",
+			id: "ysharma/Explore_llamav2_with_TGI",
 		},
 		{
-			description: "An text generation based application for conversations.",
-			id: "h2oai/h2ogpt-chatbot",
+			description: "An text generation based application to converse with Zephyr model.",
+			id: "HuggingFaceH4/zephyr-chat",
 		},
 		{
 			description: "An text generation application that combines OpenAI and Hugging Face models.",
 			id: "microsoft/HuggingGPT",
 		},
 		{
-			description: "An text generation application that uses StableLM-tuned-alpha-7b.",
-			id: "stabilityai/stablelm-tuned-alpha-chat",
-		},
-		{
-			description: "An UI that uses StableLM-tuned-alpha-7b.",
-			id: "togethercomputer/OpenChatKit",
+			description: "An chatbot to converse with a very powerful text generation model.",
+			id: "mlabonne/phixtral-chat",
 		},
 	],
 	summary:
-		"Generating text is the task of producing new text. These models can, for example, fill in incomplete text or paraphrase.",
+		"Generating text is the task of generating new text given another text. These models can, for example, fill in incomplete text or paraphrase.",
 	widgetModels: ["HuggingFaceH4/zephyr-7b-beta"],
 	youtubeId: "Vpjb1lu0MDk",
 };

package/src/tasks/text-generation/inference.ts

@@ -0,0 +1,194 @@
+/**
+ * Inference code generated from the JSON schema spec in ./spec
+ *
+ * Using src/scripts/inference-codegen
+ */
+
+/**
+ * Inputs for Text Generation inference
+ */
+export interface TextGenerationInput {
+	/**
+	 * The text to initialize generation with
+	 */
+	inputs: string;
+	/**
+	 * Additional inference parameters
+	 */
+	parameters?: TextGenerationParameters;
+	[property: string]: unknown;
+}
+
+/**
+ * Additional inference parameters
+ *
+ * Additional inference parameters for Text Generation
+ */
+export interface TextGenerationParameters {
+	/**
+	 * The number of sampling queries to run. Only the best one (in terms of total logprob) will
+	 * be returned.
+	 */
+	best_of?: number;
+	/**
+	 * Whether or not to output decoder input details
+	 */
+	decoder_input_details?: boolean;
+	/**
+	 * Whether or not to output details
+	 */
+	details?: boolean;
+	/**
+	 * Whether to use logits sampling instead of greedy decoding when generating new tokens.
+	 */
+	do_sample?: boolean;
+	/**
+	 * The maximum number of tokens to generate.
+	 */
+	max_new_tokens?: number;
+	/**
+	 * The parameter for repetition penalty. A value of 1.0 means no penalty. See [this
+	 * paper](https://hf.co/papers/1909.05858) for more details.
+	 */
+	repetition_penalty?: number;
+	/**
+	 * Whether to prepend the prompt to the generated text.
+	 */
+	return_full_text?: boolean;
+	/**
+	 * The random sampling seed.
+	 */
+	seed?: number;
+	/**
+	 * Stop generating tokens if a member of `stop_sequences` is generated.
+	 */
+	stop_sequences?: string[];
+	/**
+	 * The value used to modulate the logits distribution.
+	 */
+	temperature?: number;
+	/**
+	 * The number of highest probability vocabulary tokens to keep for top-k-filtering.
+	 */
+	top_k?: number;
+	/**
+	 * If set to < 1, only the smallest set of most probable tokens with probabilities that add
+	 * up to `top_p` or higher are kept for generation.
+	 */
+	top_p?: number;
+	/**
+	 * Truncate input tokens to the given size.
+	 */
+	truncate?: number;
+	/**
+	 * Typical Decoding mass. See [Typical Decoding for Natural Language
+	 * Generation](https://hf.co/papers/2202.00666) for more information
+	 */
+	typical_p?: number;
+	/**
+	 * Watermarking with [A Watermark for Large Language Models](https://hf.co/papers/2301.10226)
+	 */
+	watermark?: boolean;
+	[property: string]: unknown;
+}
+
+/**
+ * Outputs for Text Generation inference
+ */
+export interface TextGenerationOutput {
+	/**
+	 * When enabled, details about the generation
+	 */
+	details?: TextGenerationOutputDetails;
+	/**
+	 * The generated text
+	 */
+	generated_text: string;
+	[property: string]: unknown;
+}
+
+/**
+ * When enabled, details about the generation
+ */
+export interface TextGenerationOutputDetails {
+	/**
+	 * Details about additional sequences when best_of is provided
+	 */
+	best_of_sequences?: TextGenerationSequenceDetails[];
+	/**
+	 * The reason why the generation was stopped.
+	 */
+	finish_reason: FinishReason;
+	/**
+	 * The number of generated tokens
+	 */
+	generated_tokens: number;
+	prefill: PrefillToken[];
+	/**
+	 * The random seed used for generation
+	 */
+	seed?: number;
+	/**
+	 * The generated tokens and associated details
+	 */
+	tokens: Token[];
+	[property: string]: unknown;
+}
+
+export interface TextGenerationSequenceDetails {
+	/**
+	 * The reason why the generation was stopped.
+	 */
+	finish_reason: FinishReason;
+	/**
+	 * The generated text
+	 */
+	generated_text: number;
+	/**
+	 * The number of generated tokens
+	 */
+	generated_tokens: number;
+	prefill: PrefillToken[];
+	/**
+	 * The random seed used for generation
+	 */
+	seed?: number;
+	/**
+	 * The generated tokens and associated details
+	 */
+	tokens: Token[];
+	[property: string]: unknown;
+}
+
+/**
+ * The generated sequence reached the maximum allowed length
+ *
+ * The model generated an end-of-sentence (EOS) token
+ *
+ * One of the sequence in stop_sequences was generated
+ */
+export type FinishReason = "length" | "eos_token" | "stop_sequence";
+
+export interface PrefillToken {
+	id: number;
+	logprob: number;
+	/**
+	 * The text associated with that token
+	 */
+	text: string;
+	[property: string]: unknown;
+}
+
+export interface Token {
+	id: number;
+	logprob: number;
+	/**
+	 * Whether or not that token is a special one
+	 */
+	special: boolean;
+	/**
+	 * The text associated with that token
+	 */
+	text: string;
+	[property: string]: unknown;
+}

package/src/tasks/text-generation/spec/input.json

@@ -0,0 +1,90 @@
+{
+	"$id": "/inference/schemas/text-generation/input.json",
+	"$schema": "http://json-schema.org/draft-06/schema#",
+	"description": "Inputs for Text Generation inference",
+	"title": "TextGenerationInput",
+	"type": "object",
+	"properties": {
+		"inputs": {
+			"description": "The text to initialize generation with",
+			"type": "string"
+		},
+		"parameters": {
+			"description": "Additional inference parameters",
+			"$ref": "#/$defs/TextGenerationParameters"
+		}
+	},
+	"$defs": {
+		"TextGenerationParameters": {
+			"title": "TextGenerationParameters",
+			"description": "Additional inference parameters for Text Generation",
+			"type": "object",
+			"properties": {
+				"best_of": {
+					"type": "integer",
+					"description": "The number of sampling queries to run. Only the best one (in terms of total logprob) will be returned."
+				},
+				"decoder_input_details": {
+					"type": "boolean",
+					"description": "Whether or not to output decoder input details"
+				},
+				"details": {
+					"type": "boolean",
+					"description": "Whether or not to output details"
+				},
+				"do_sample": {
+					"type": "boolean",
+					"description": "Whether to use logits sampling instead of greedy decoding when generating new tokens."
+				},
+				"max_new_tokens": {
+					"type": "integer",
+					"description": "The maximum number of tokens to generate."
+				},
+				"repetition_penalty": {
+					"type": "number",
+					"description": "The parameter for repetition penalty. A value of 1.0 means no penalty. See [this paper](https://hf.co/papers/1909.05858) for more details."
+				},
+				"return_full_text": {
+					"type": "boolean",
+					"description": "Whether to prepend the prompt to the generated text."
+				},
+				"seed": {
+					"type": "integer",
+					"description": "The random sampling seed."
+				},
+				"stop_sequences": {
+					"type": "array",
+					"items": {
+						"type": "string"
+					},
+					"description": "Stop generating tokens if a member of `stop_sequences` is generated."
+				},
+				"temperature": {
+					"type": "number",
+					"description": "The value used to modulate the logits distribution."
+				},
+				"top_k": {
+					"type": "integer",
+					"description": "The number of highest probability vocabulary tokens to keep for top-k-filtering."
+				},
+				"top_p": {
+					"type": "number",
+					"description": "If set to < 1, only the smallest set of most probable tokens with probabilities that add up to `top_p` or higher are kept for generation."
+				},
+				"truncate": {
+					"type": "integer",
+					"description": "Truncate input tokens to the given size."
+				},
+				"typical_p": {
+					"type": "number",
+					"description": "Typical Decoding mass. See [Typical Decoding for Natural Language Generation](https://hf.co/papers/2202.00666) for more information"
+				},
+				"watermark": {
+					"type": "boolean",
+					"description": "Watermarking with [A Watermark for Large Language Models](https://hf.co/papers/2301.10226)"
+				}
+			}
+		}
+	},
+	"required": ["inputs"]
+}

package/src/tasks/text-generation/spec/output.json

@@ -0,0 +1,120 @@
+{
+	"$id": "/inference/schemas/text-generation/output.json",
+	"$schema": "http://json-schema.org/draft-06/schema#",
+	"description": "Outputs for Text Generation inference",
+	"title": "TextGenerationOutput",
+	"type": "object",
+	"properties": {
+		"generated_text": {
+			"type": "string",
+			"description": "The generated text"
+		},
+		"details": {
+			"description": "When enabled, details about the generation",
+			"title": "TextGenerationOutputDetails",
+			"allOf": [
+				{ "$ref": "#/$defs/SequenceDetails" },
+				{
+					"type": "object",
+					"properties": {
+						"best_of_sequences": {
+							"type": "array",
+							"description": "Details about additional sequences when best_of is provided",
+							"items": {
+								"allOf": [
+									{ "$ref": "#/$defs/SequenceDetails" },
+									{
+										"type": "object",
+										"properties": {
+											"generated_text": {
+												"type": "integer",
+												"description": "The generated text"
+											}
+										},
+										"required": ["generated_text"]
+									}
+								]
+							}
+						}
+					}
+				}
+			]
+		}
+	},
+	"required": ["generated_text"],
+
+	"$defs": {
+		"Token": {
+			"type": "object",
+			"title": "Token",
+			"properties": {
+				"id": {
+					"type": "integer"
+				},
+				"logprob": {
+					"type": "number"
+				},
+				"special": {
+					"type": "boolean",
+					"description": "Whether or not that token is a special one"
+				},
+				"text": {
+					"type": "string",
+					"description": "The text associated with that token"
+				}
+			},
+			"required": ["id", "logprob", "special", "text"]
+		},
+		"SequenceDetails": {
+			"type": "object",
+			"title": "TextGenerationSequenceDetails",
+			"properties": {
+				"finish_reason": {
+					"type": "string",
+					"description": "The reason why the generation was stopped.",
+					"oneOf": [
+						{ "const": "length", "description": "The generated sequence reached the maximum allowed length" },
+						{ "const": "eos_token", "description": "The model generated an end-of-sentence (EOS) token" },
+						{ "const": "stop_sequence", "description": "One of the sequence in stop_sequences was generated" }
+					]
+				},
+				"generated_tokens": {
+					"type": "integer",
+					"description": "The number of generated tokens"
+				},
+				"prefill": {
+					"type": "array",
+					"items": {
+						"title": "PrefillToken",
+						"type": "object",
+						"properties": {
+							"id": {
+								"type": "integer"
+							},
+							"logprob": {
+								"type": "number"
+							},
+							"text": {
+								"type": "string",
+								"description": "The text associated with that token"
+							}
+						},
+						"required": ["id", "logprob", "text"]
+					}
+				},
+				"seed": {
+					"type": "integer",
+					"description": "The random seed used for generation"
+				},
+				"tokens": {
+					"type": "array",
+					"description": "The generated tokens and associated details",
+					"items": {
+						"$ref": "#/$defs/Token"
+					}
+				}
+			},
+			"required": ["finish_reason", "generated_tokens", "prefill", "tokens"]
+		}
+	}
+}

package/src/tasks/text-to-audio/inference.ts

@@ -0,0 +1,143 @@
+/**
+ * Inference code generated from the JSON schema spec in ./spec
+ *
+ * Using src/scripts/inference-codegen
+ */
+
+/**
+ * Inputs for Text To Audio inference
+ */
+export interface TextToAudioInput {
+	/**
+	 * The input text data
+	 */
+	inputs: string;
+	/**
+	 * Additional inference parameters
+	 */
+	parameters?: TextToAudioParameters;
+	[property: string]: unknown;
+}
+
+/**
+ * Additional inference parameters
+ *
+ * Additional inference parameters for Text To Audio
+ */
+export interface TextToAudioParameters {
+	/**
+	 * Parametrization of the text generation process
+	 */
+	generate?: GenerationParameters;
+	[property: string]: unknown;
+}
+
+/**
+ * Parametrization of the text generation process
+ *
+ * Ad-hoc parametrization of the text generation process
+ */
+export interface GenerationParameters {
+	/**
+	 * Whether to use sampling instead of greedy decoding when generating new tokens.
+	 */
+	do_sample?: boolean;
+	/**
+	 * Controls the stopping condition for beam-based methods.
+	 */
+	early_stopping?: EarlyStoppingUnion;
+	/**
+	 * If set to float strictly between 0 and 1, only tokens with a conditional probability
+	 * greater than epsilon_cutoff will be sampled. In the paper, suggested values range from
+	 * 3e-4 to 9e-4, depending on the size of the model. See [Truncation Sampling as Language
+	 * Model Desmoothing](https://hf.co/papers/2210.15191) for more details.
+	 */
+	epsilon_cutoff?: number;
+	/**
+	 * Eta sampling is a hybrid of locally typical sampling and epsilon sampling. If set to
+	 * float strictly between 0 and 1, a token is only considered if it is greater than either
+	 * eta_cutoff or sqrt(eta_cutoff) * exp(-entropy(softmax(next_token_logits))). The latter
+	 * term is intuitively the expected next token probability, scaled by sqrt(eta_cutoff). In
+	 * the paper, suggested values range from 3e-4 to 2e-3, depending on the size of the model.
+	 * See [Truncation Sampling as Language Model Desmoothing](https://hf.co/papers/2210.15191)
+	 * for more details.
+	 */
+	eta_cutoff?: number;
+	/**
+	 * The maximum length (in tokens) of the generated text, including the input.
+	 */
+	max_length?: number;
+	/**
+	 * The maximum number of tokens to generate. Takes precedence over maxLength.
+	 */
+	max_new_tokens?: number;
+	/**
+	 * The minimum length (in tokens) of the generated text, including the input.
+	 */
+	min_length?: number;
+	/**
+	 * The minimum number of tokens to generate. Takes precedence over maxLength.
+	 */
+	min_new_tokens?: number;
+	/**
+	 * Number of groups to divide num_beams into in order to ensure diversity among different
+	 * groups of beams. See [this paper](https://hf.co/papers/1610.02424) for more details.
+	 */
+	num_beam_groups?: number;
+	/**
+	 * Number of beams to use for beam search.
+	 */
+	num_beams?: number;
+	/**
+	 * The value balances the model confidence and the degeneration penalty in contrastive
+	 * search decoding.
+	 */
+	penalty_alpha?: number;
+	/**
+	 * The value used to modulate the next token probabilities.
+	 */
+	temperature?: number;
+	/**
+	 * The number of highest probability vocabulary tokens to keep for top-k-filtering.
+	 */
+	top_k?: number;
+	/**
+	 * If set to float < 1, only the smallest set of most probable tokens with probabilities
+	 * that add up to top_p or higher are kept for generation.
+	 */
+	top_p?: number;
+	/**
+	 * Local typicality measures how similar the conditional probability of predicting a target
+	 * token next is to the expected conditional probability of predicting a random token next,
+	 * given the partial text already generated. If set to float < 1, the smallest set of the
+	 * most locally typical tokens with probabilities that add up to typical_p or higher are
+	 * kept for generation. See [this paper](https://hf.co/papers/2202.00666) for more details.
+	 */
+	typical_p?: number;
+	/**
+	 * Whether the model should use the past last key/values attentions to speed up decoding
+	 */
+	use_cache?: boolean;
+	[property: string]: unknown;
+}
+
+/**
+ * Controls the stopping condition for beam-based methods.
+ */
+export type EarlyStoppingUnion = boolean | "never";
+
+/**
+ * Outputs of inference for the Text To Audio task
+ */
+export interface TextToAudioOutput {
+	/**
+	 * The generated audio waveform.
+	 */
+	audio: unknown;
+	samplingRate: unknown;
+	/**
+	 * The sampling rate of the generated audio waveform.
+	 */
+	sampling_rate?: number;
+	[property: string]: unknown;
+}

package/src/tasks/text-to-audio/spec/input.json

@@ -0,0 +1,31 @@
+{
+	"$id": "/inference/schemas/text-to-audio/input.json",
+	"$schema": "http://json-schema.org/draft-06/schema#",
+	"description": "Inputs for Text To Audio inference",
+	"title": "TextToAudioInput",
+	"type": "object",
+	"properties": {
+		"inputs": {
+			"description": "The input text data",
+			"type": "string"
+		},
+		"parameters": {
+			"description": "Additional inference parameters",
+			"$ref": "#/$defs/TextToAudioParameters"
+		}
+	},
+	"$defs": {
+		"TextToAudioParameters": {
+			"title": "TextToAudioParameters",
+			"description": "Additional inference parameters for Text To Audio",
+			"type": "object",
+			"properties": {
+				"generate": {
+					"description": "Parametrization of the text generation process",
+					"$ref": "/inference/schemas/common-definitions.json#/definitions/GenerationParameters"
+				}
+			}
+		}
+	},
+	"required": ["inputs"]
+}