GizStackGizStack
View .mdOpen in ChatGPTOpen in Claude

gizai/models/ardy/text

ARDY's text encoder, LLM2Vec of Llama 3 8B Instruct (McGill-NLP's MNTP and supervised LoRA adapters merged at conversion) on gizai/llm. The prompt is a user turn of Llama 3's chat format, as LLM2Vec trained its supervised adapter (and as ARDY reads its text from local folders; from the hub's names LLM2Vec leaves the turn out). Every token sees every other (a bidirectional pass), and the prompt's vector is the mean of the rows of its words and its end of turn after the final norm.

gizlang
import * as Text from "gizai/models/ardy/text";

Types

Text

gizlang
type Text = { model: LlmState.Model; known: Map<string, Vec<f32>>; }

The text encoder and the vectors of the prompts it has read.

  • model: M53.Model
  • known: _gizlang_sdk_collections.Map<string, Vec<f32>>

Functions

folderIn

gizlang
function folderIn(hub: Hub.Hub): string throws string

The converted encoder in hub: downloaded and converted once.

Effects: fail, read, write, console, io, alloc.

load

gizlang
function load(folder: string, gpu: Option<Gpu>, after: i64 = 0): Text throws string

The encoder of folder (folderIn's) on gpu (none: the host); after: the bytes the models it serves load after it (encoderModel).

Effects: fail, read, write, console, io, alloc.

Methods

Text.vector

gizlang
function Text.vector(self: mut Text, prompt: string): Vec<f32> throws string

The vector of prompt (TEXT values), read once.

Effects: fail, read, write, console, io, alloc.