Nethena-20b-Glue-LORA

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

<img src="https://raw.githubusercontent.com/OpenAccess-AI-Collective/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>

This model is a fine-tuned version of NeverSleep/Nethena-20B on a private dataset. It achieves the following results on the evaluation set:

• Loss: 1.3864

Model description

athirdpath/Nethena-20b-Glued-LORA is a 128 rank LORA for RP, trained on NeverSleep/Nethena-20B. It is unalligned and NSFW-oriented.

This is a test, exploring the effects of "gluing" the components of the 20b model together to reduce the iconic word replacement errors, increase lucidity, and improve recall.

Training and evaluation data

The private ~500k token dataset used to train the LORA was Alpaca formatted and focused on 4 primary categories:

- Medical texts (on psychology, reproductive organs, anatomy, and pregnancy). These are formatted so the model, in character as a doctor or therapist, answers a patient's question in short to medium form.
- Excerpts from short stories and novellas (erotic and romantic) centered around both realistic and fantastic situations, covering several fetishes as well. These are sliced into ~2048 token chunks, and these long-form responses are all tied to the command “Enter narrator mode.” in the instructions.
- A selection from PIPPA, using a wide keyword search for tokens associated with low quality human or AI data to remove those responses, then a positive search was done for words and phrases associated with a higher reading level. These are converted to Alpaca with “Enter RP mode.” in all the instruction fields.
- ~18k tokens of GPT-4 generated data on role-playing from various characters’ perspectives, focusing on different situations and emotions. Includes many multi-turn conversations.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 5
• total_train_batch_size: 20
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• num_epochs: 3

Training results

Framework versions

• Transformers 4.34.1
• Pytorch 2.1.0+cu118
• Datasets 2.14.6
• Tokenizers 0.14.1