Dans-Archive/Dans-PileOfSets-Mk1-llama-13b-merged
Description:
This is a llama 13b model merge of the LoRA with the same name.
Objective for this project:
To create a model that upholds a logical thread, regardless of whether the output is verbose or concise. Training has been performed on a version of the pile of sets, reduced to 40% of its original size, to expedite training iterations. I personally utilize this model as an aid for storytelling and writing. While it serves this purpose adequately, I still perceive this version as a prototype.
Prompt format:
Stanford Alpaca
The prompt should start on a new line after "### Response:"
- For examples with a non-empty input field:
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:
{instruction}
### Input:
{input}
### Response:- For examples with an empty input field:
Below is an instruction that describes a task. Write a response that appropriately completes the request.
### Instruction:
{instruction}
### Response:Perplexity Benchmarks:
- wikitext: 4.66796875
Training information:
- 2 Epochs
- 64 / 32 R / A
- 1024 Cutoff
- 19 hours on an A6000
Data used in training:
All cleaned and scrubbed in various ways then culled to various degrees.
- Camel biology, physics, chemistry, math, and AI society
- Alpaca evol instruct
- GPTeacher Instruct
- Alpaca GPT4
- Dolly Databricks
Plans for the future, a brief overview:
- Pivot to a conversational format going forward
- Train another 13b LoRA against the entirety of my pile of sets rather than just a portion of it for Mk2
- Train 30b on the Mk2 pile of sets
- Expand the story generation capabilities and likely more for Mk3
Model used for training and other information:
https://huggingface.co/PocketDoc/llama-13b-gptq-4bit-128g
Merge model: https://huggingface.co/huggyllama/llama-13b
Disclaimer:
It has not been aligned and no warranty is given for the quality or safety of its outputs.