Built with Axolotl

storage/context

This model is a fine-tuned version of mistralai/Mistral-7B-v0.1 on the mhenrichsen/context-aware-splits-english dataset. It achieves the following results on the evaluation set:

  • Loss: 0.0253

Model description

  • This model is used to split texts in a context aware way. Used for RAG applications.
  • This model is an adapter for Mistral 7b. It uses the Alpaca format:
Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
Your task is to segment text into smaller blocks. Split the text where it makes sense and be vary of the context. The ideal split should be close to {WORD_COUNT} words.

### Input:
Q: Information/File Manager I'm looking for a file manager application which helps to organize a large amount of movies, pictures, music, text documents, databases, audio-books and ebooks. Right now I only use the Finder which doesn't work well, because I really need a function to put single files into multiple categories. Simply using the file system for this creates a confusing nesting of files. A: Depending on the number of categories you require to handle, you could always use a combination of the finder with the built in label functionality, thus a movie can be held in one area (movies directory, for example), but "tagged" as something else. Using smart directories and saved searches you can view your files by a combination of the attributes (location, label, media type) to create custom views. All without purchasing software. Cheap and cheerful, but may be suitable to your needs. A: Maybe use a file manager that supports Open Meta. Or use symbolic links for organizing all your media files. Or even use hardlinked files if you dare.

### Response:

Response:

{'splits': ["Q: Information/File Manager I'm looking for a file manager application which helps to organize a large amount of movies, pictures, music, text documents, databases, audio-books and ebooks. Right now I only use the Finder which doesn't work well, because I really need a function to put single files into multiple categories. Simply using the file system for this creates a confusing nesting of files.", 'A: Depending on the number of categories you require to handle, you could always use a combination of the finder with the built in label functionality, thus a movie can be held in one area (movies directory, for example), but "tagged" as something else. Using smart directories and saved searches you can view your files by a combination of the attributes (location, label, media type) to create custom views. All without purchasing software. Cheap and cheerful, but may be suitable to your needs.', 'A: Maybe use a file manager that supports Open Meta. Or use symbolic links for organizing all your media files. Or even use hardlinked files if you dare.'], 'topic': 'Discussion on file manager applications for organizing large amount of media files.'}

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 0.0002
  • train_batch_size: 4
  • eval_batch_size: 4
  • seed: 42
  • gradient_accumulation_steps: 8
  • total_train_batch_size: 32
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: cosine
  • lr_scheduler_warmup_steps: 10
  • num_epochs: 1

Training results

Training Loss Epoch Step Validation Loss
0.1557 0.01 1 0.1552
0.0936 0.05 6 0.0695
0.0447 0.1 12 0.0413
0.0347 0.16 18 0.0357
0.0314 0.21 24 0.0324
0.0306 0.26 30 0.0309
0.0276 0.31 36 0.0294
0.028 0.36 42 0.0284
0.0307 0.41 48 0.0281
0.0276 0.47 54 0.0274
0.0251 0.52 60 0.0267
0.0244 0.57 66 0.0269
0.0268 0.62 72 0.0263
0.0249 0.67 78 0.0262
0.0252 0.73 84 0.0258
0.0259 0.78 90 0.0257
0.0241 0.83 96 0.0255
0.0241 0.88 102 0.0254
0.0253 0.93 108 0.0254
0.0234 0.98 114 0.0253

Framework versions

  • Transformers 4.35.2
  • Pytorch 2.0.1+cu118
  • Datasets 2.15.0
  • Tokenizers 0.15.0
Downloads last month
15
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Model tree for mhenrichsen/context-aware-splitter-mistral-eng

Quantized
(168)
this model

Dataset used to train mhenrichsen/context-aware-splitter-mistral-eng

Space using mhenrichsen/context-aware-splitter-mistral-eng 1