exo/lm-dataset-preparer
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Readme;

Browse Source
This commit is contained in:
extio-de
2024-10-30 00:22:54 +01:00
parent dca4aa0a29
commit 2087405c64
1 changed files with 9 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
README.md
View File

@@ -1,26 +1,19 @@
# lm-dataset-preparer # lm-dataset-preparer
**Program Name:** Text Processing and Translation Program This program is designed for processing and rewriting text files using a language model.
**Purpose:** The program is designed to process and translate text files, generating new paragraphs and converting them into various formats. It utilizes a client to make API calls to a model, applies text normalization and transformation techniques, and writes the results to new files. Here's a breakdown of its functionality:
**Key Features:** 1. **Text Preparation:** The program utilizes several classes to prepare the text for processing. It can translate text files to English, normalize text, split text into paragraphs, and convert paragraphs into JSON format.
1. **Text Translation:** Translates text files from various languages to English using a client-provided completion API. 2. **Contextual Prompt Generation:** For each paragraph, the program generates contextual prompts using an external AI service. These prompts help guide the language model in understanding the context of the paragraph.
2. **Model-based Rewriting:** Generates new paragraphs based on input text and model categories, applying multiple models and enhancements.
3. **Text Normalization:** Removes excessive newlines, spaces, and tabs, and replaces multiple consecutive newlines with a single newline.
4. **Paragraph Splitting:** Splits text into paragraphs based on a specified chunk size and tolerance range.
5. **Directory Transformation:** Applies a function to each file in a directory in parallel using a fixed thread pool.
6. **File Conversion:** Converts text files into JSONL (JSON Lines) format, splitting them into paragraphs and chunks based on provided normalization and variation values.
**Components:** 3. **Text Rewriting:** The program leverages a language model to rewrite text paragraphs. It supports two types of rewriting: "improvement" and "enhancement," each with its own set of prompts. Multiple enhancement variations are possible by randomly selecting from a list of prompts.
1. A Spring component that translates text files to English. 4. **Output:** The rewritten text, along with the generated prompts, is saved in a JSONL file.
2. A Spring component that rewrites text files by generating new paragraphs based on input text and model categories.
3. A Spring component that converts text files into JSONL format.
4. A utility class with methods for text normalization, paragraph splitting, and directory transformation.
Overall, the program is designed to efficiently process and translate text files, applying various techniques to generate new paragraphs and convert them into different formats. **Purpose:**
*[requests=5, requestDuration=PT14.64756857S, inTokens=3966, outTokens=801, tps=57.214285714285715]* The overall purpose of this program is to facilitate the rewriting and improvement of text files using a language model. The contextual prompts and different rewriting modes allow for flexible and nuanced text manipulation. This could be useful for tasks such as fine-tuning of large language models (LLMs).
*[gemma-2-27b-it-Q4_K_L.gguf; requests=6, requestDuration=PT1M36.661188325S, inTokens=6473, outTokens=506, tps=5.270833333333333]*