Triangle104/Pygmalion-3-12B-Q4_K_S-GGUF

This model was converted to GGUF format from PygmalionAI/Pygmalion-3-12B using llama.cpp via the ggml.ai's GGUF-my-repo space. Refer to the original model card for more details on the model.

---

Model Details

It's been a long road fraught with delays, technical issues and us banging our heads against the wall, but we're glad to say that we've returned to open-source roleplaying with our newest model, Pygmalion-3. We've taken Mistral's Nemo base model and fed it hundreds of millions of tokens of conversations, creative writing and instructions to create a model dedicated towards roleplaying that we hope fulfills your expectations.

As part of our open-source roots and promises to those who have been with us since the beginning, we release this model under the permissive Apache 2.0 license, allowing anyone to use and develop upon our work for everybody in the local models community.

Prompting

We've switched over to standard ChatML format for both convenience and to allow easier merging with other ChatML-based models. Pygmalion-3 supports the same "Enter X mode" as our previous Pygmalion-2 model, but experimentation with the system prompt is encouraged so that you can find what is best for you.

Please note that some strange issues have been reported with the <|im_end|> token. It is highly recommended to add a custom token ban on the phrase "<|im_end|>" and "<" in general. We apologize for the inconvenience.

Prompting example

<|im_start|>system Enter roleplay mode. You shall reply to {{user}} while staying in character. Your responses must be detailed, creative, immersive, and drive the scenario forward. You will follow {{char}}'s persona.<|im_end|> <|im_start|>user {{user}}: Good evening!<|im_end|> <|im_start|>assistant {{char}}: It's three in the morning, man.<|im_end|>

Note that {{user}} and {{char}} are placeholders.

Dataset

We've gathered a large collection of instructions and roleplaying totaling hundreds of millions of tokens, including our PIPPA dataset and roleplaying forums.

Limitations and biases

The intended use-case for this model is fictional writing for entertainment purposes. Any other sort of usage is out of scope.

As such, it was not fine-tuned to be safe and harmless: the base model and this fine-tune have been trained on data known to contain profanity and texts that are lewd or otherwise offensive. It may produce socially unacceptable or undesirable text, even if the prompt itself does not include anything explicitly offensive. Outputs might often be factually wrong or misleading.

Training Specifications

We trained our model as a rank-32 LoRA adapter with one epoch over our data using 8x NVIDIA A40 GPUs. For this run, we employed a learning rate of 2e-4 and a total batch size across all GPUs of 24. A cosine learning rate scheduler was used with a 100 step warmup. DeepSpeed ZeRO was used to successfully get memory usage down.

Acknowledgements

This project could not have been done without the compute support of Hive Digital Technologies and the Axolotl training software.

We'd like to extensively thank lemonilia for their wonderful help in compiling roleplay forum data.

And most of all, we dedicate this model to our great community, who've stuck with us through everything until now. Sincerely, thank you so much. We hope you enjoy our work to the fullest and we promise more is on the way soon.

---

Use with llama.cpp

Install llama.cpp through brew (works on Mac and Linux)

brew install llama.cpp

Invoke the llama.cpp server or the CLI.

CLI:

llama-cli --hf-repo Triangle104/Pygmalion-3-12B-Q4_K_S-GGUF --hf-file pygmalion-3-12b-q4_k_s.gguf -p "The meaning to life and the universe is"

Server:

llama-server --hf-repo Triangle104/Pygmalion-3-12B-Q4_K_S-GGUF --hf-file pygmalion-3-12b-q4_k_s.gguf -c 2048

Note: You can also use this checkpoint directly through the usage steps listed in the Llama.cpp repo as well.

Step 1: Clone llama.cpp from GitHub.

git clone https://github.com/ggerganov/llama.cpp

Step 2: Move into the llama.cpp folder and build it with LLAMA_CURL=1 flag along with other hardware-specific flags (for ex: LLAMA_CUDA=1 for Nvidia GPUs on Linux).

cd llama.cpp && LLAMA_CURL=1 make

Step 3: Run inference through the main binary.

./llama-cli --hf-repo Triangle104/Pygmalion-3-12B-Q4_K_S-GGUF --hf-file pygmalion-3-12b-q4_k_s.gguf -p "The meaning to life and the universe is"

or

./llama-server --hf-repo Triangle104/Pygmalion-3-12B-Q4_K_S-GGUF --hf-file pygmalion-3-12b-q4_k_s.gguf -c 2048