crestf411/L3.1-nemotron-sunfall-v0.7.0

Sunfall (2024-10-28) v0.7.0 trained directly against, and merged with Nemotron 70B Instruct.

It also contains samples from Antracite.Org datasets. See bottom for details.

Significant revamping of the dataset metadata generation process, resulting in higher quality dataset overall. The "Diamond Law" experiment has been removed as it didn't seem to affect the model output enough to warrant set up complexity.

Recommended starting point:

• Temperature: 1
• MinP: 0.05~0.1
• DRY: 0.8 1.75 2 0

At early context, I recommend keeping XTC disabled. Once you hit higher context sizes (10k+), enabling XTC at 0.1 / 0.5 seems to significantly improve the output, but YMMV. If the output drones on and is uninspiring, XTC can be extremely effective.

General heuristic:

• Lots of slop? Temperature is too low. Raise it, or enable XTC. For early context, temp bump is probably preferred.
• Is the model making mistakes about subtle or obvious details in the scene? Temperature is too high, OR XTC is enabled and/or XTC settings are too high. Lower temp and/or disable XTC.

Mergers/fine-tuners: there is a LoRA of this model. Consider merging that instead of merging this model.

This model has been trained on context that mimics that of Silly Tavern's "Llama 3 Instruct" preset, with "Always add character's name to prompt" checked.

The model has also been trained to do interactive storywriting. You may steer the model towards specific content by "responding" to the model like so:

Continue writing adhering to the following scenario: (things you want to happen next)

Additional inclusions (random sampled sub-set, cursorily quality-checked) from:

• Gryphe/Sonnet3.5-Charcard-Roleplay
• anthracite-org/c2_logs_32k_mistral-v3_v1.2_no_system
• anthracite-org/kalo-opus-instruct-22k-no-refusal-no-system
• anthracite-org/kalo-opus-instruct-3k-filtered-no-system
• anthracite-org/nopm_claude_writing_fixed

As such, the dataset is not 100% slop free, but this addition likely helps the model be a better roleplayer. At some point, I intend to clean up and release the samples, deslopped.

Note on training:

The training was done using Fine-Tuning with Very Large Dropout (h/t https://huggingface.co/Envoid/Llama-3.05-NT-Storybreaker-Ministral-70B for the idea) with a LoRA dropout of 0.5 and a constant learning rate of 4e-6. In addition, the model seemed to retain more of Nemotron's smartness by halving the alpha, which is how this merge (and the LoRA adapter configuration) is set up. (The LoRA was trained with alpha=64, and merged with alpha set to 32.)