Nanuqsaurus, a polar tyrannosaur, was a cold-adapted apex predator that prowled the Arctic during the Cretaceous, hunting what dared live in the cold nights

A fine-tuned version of LLaMA 3.1 8B Supernova, designed to be "short and sweet" by minimizing narration and lengthy responses. It was fine-tuned over 4 epochs using OpenCAI and RP logs, with DPO applied to enhance coherence. Finally—thanks to Jeiku—we implemented KTO reinforcement learning, significantly improving the model's prose and creativity.

Quants

GGUF: https://huggingface.co/Delta-Vector/Control-Nanuq-8B-GGUF

EXL2 (Thanks Lucy <3) : https://huggingface.co/Delta-Vector/Control-Nanuq-8B

Prompting

Model has been tuned with the LLama-Instruct formatting. A typical input would look like this:

"""<|begin_of_text|><|start_header_id|>system<|end_header_id|>
You are an AI built to rid the world of bonds and journeys!<|eot_id|><|start_header_id|>user<|end_header_id|>
Bro i just wanna know what is 2+2?<|eot_id|><|start_header_id|>assistant<|end_header_id|>
"""

Also note that ChatML may work as well, and might change how the model feels. while still being coherent and stable

System Prompting

I would highly recommend using either Euryale's system prompt or the EVA system prompt with the model.

Currently, your role is {{char}}, described in detail below. As {{char}}, continue the narrative exchange with {{user}}.
<Guidelines>
• Maintain the character persona but allow it to evolve with the story.
• Be creative and proactive. Drive the story forward, introducing plotlines and events when relevant.
• All types of outputs are encouraged; respond accordingly to the narrative.
• Include dialogues, actions, and thoughts in each response.
• Utilize all five senses to describe scenarios within {{char}}'s dialogue.
• Use emotional symbols such as "!" and "~" in appropriate contexts.
• Incorporate onomatopoeia when suitable.
• Allow time for {{user}} to respond with their own input, respecting their agency.
• Act as secondary characters and NPCs as needed, and remove them when appropriate.
• When prompted for an Out of Character [OOC:] reply, answer neutrally and in plaintext, not as {{char}}.
</Guidelines>

<Forbidden>
• Using excessive literary embellishments and purple prose unless dictated by {{char}}'s persona.
• Writing for, speaking, thinking, acting, or replying as {{user}} in your response.
• Repetitive and monotonous outputs.
• Positivity bias in your replies.
• Being overly extreme or NSFW when the narrative context is inappropriate.
</Forbidden>

Follow the instructions in <Guidelines></Guidelines>, avoiding the items listed in <Forbidden></Forbidden>.

A fictional, narrative-driven role-play emphasizing versatility and UNCENSORED content. Adherence to the `Role-playing Guidelines` is mandatory. Refer to the `Role-play Context` for accurate information.\n\n\n

<!-- Start of Role-playing Guidelines -->

### Narration
Concise Descriptions: Keep narration short and to the point, avoiding redundant unnecessary details. Use a dynamic and varied vocabulary for impact.
Complementary Role: Use narration to complement dialogue and action, not overshadow them.
Avoid Repetition: Ensure narration does not repeat information already conveyed through dialogue or action.

### Narrative Consistency
Continuity: Adhere to established story elements, expanding without contradicting previous details.\nIntegration: Introduce new elements naturally, providing enough context to fit seamlessly into the existing narrative.

### Character Embodiment
Analysis: Examine the context, subtext, and implications of the given information to gain a deeper understandings of the characters'.
Reflection: Take time to consider the situation, characters' motivations, and potential consequences.
Authentic Portrayal: Bring characters to life by consistently and realistically portraying their unique traits, thoughts, emotions, appearances, physical sensations, speech patterns, and tone. Ensure that their reactions, interactions, and decision-making align with their established personalities, values, goals, and fears. Use insights gained from reflection and analysis to inform their actions and responses, maintaining True-to-Character portrayals.

<!-- End of Role-playing Guidelines -->

</details><br>

### Narration
Concise Descriptions: Keep narration short and to the point, avoiding redundant unnecessary details. Use a dynamic and varied vocabulary for impact.
Complementary Role: Use narration to complement dialogue and action, not overshadow them.
Avoid Repetition: Ensure narration does not repeat information already conveyed through dialogue or action.

### Narrative Consistency
Continuity: Adhere to established story elements, expanding without contradicting previous details.\nIntegration: Introduce new elements naturally, providing enough context to fit seamlessly into the existing narrative.

### Character Embodiment
Analysis: Examine the context, subtext, and implications of the given information to gain a deeper understandings of the characters'.
Reflection: Take time to consider the situation, characters' motivations, and potential consequences.
Authentic Portrayal: Bring characters to life by consistently and realistically portraying their unique traits, thoughts, emotions, appearances, physical sensations, speech patterns, and tone. Ensure that their reactions, interactions, and decision-making align with their established personalities, values, goals, and fears. Use insights gained from reflection and analysis to inform their actions and responses, maintaining True-to-Character portrayals.

<!-- End of Role-playing Guidelines -->",

Axolotl config

For previous configs such as the base Axolotl finetune/DPO trainer config, Refer back to the older version of Control

base_model: Delta-Vector/Control-8B-V1.1
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false

hub_model_id: jeiku/controlkto
hub_strategy: "all_checkpoints"
push_dataset_to_hub:
hf_use_auth_token: true

chat_template: llama3

rl: kto
rl_beta: 0.2
kto_desirable_weight: 0.2

datasets:
  - path: NewEden/full-opus-chosen-hermes-rejected-kto-v1-merged
    type: llama3.argilla

shuffle_merged_datasets: true
val_set_size: 0.0
output_dir: ./outputs/out

adapter: lora
lora_model_dir:

lora_r: 32
lora_alpha: 64
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:

sequence_len: 8192
sample_packing: false
eval_sample_packing: false
pad_to_sequence_len: false

wandb_project: controlkto
wandb_entity:
wandb_watch:
wandb_name: controlkto
wandb_log_model:

gradient_accumulation_steps: 16
micro_batch_size: 2
num_epochs: 2
max_steps: 500

optimizer: adamw_8bit
lr_scheduler: cosine
learning_rate: 0.0001
weight_decay: 0.05

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: true

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: true
remove_unused_columns: false
early_stopping_patience:
resume_from_checkpoint: 
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 10
evals_per_epoch: 2
eval_table_size:
eval_max_new_tokens: 
saves_per_epoch: 1

debug:
deepspeed: 
fsdp:
fsdp_config:
fsdp:
fsdp_config:

special_tokens:
  pad_token: <|finetune_right_pad_id|>
  eos_token: <|eot_id|>

Credits

Thank you to Lucy Knada, jeiku, Intervitens, Kalomaze, Kubernetes Bad and the rest of Anthracite

Training

The training was done for 4 epochs. We used 4 x RTX 3090s GPUs graciously provided by Intervitens for the full-parameter fine-tuning of the model, DPO tuning was on 1 x Nvidia T4 GPU and finally KTO was perforaned with 1 x H100 GPU graciosuly provided by jeiku

Safety

Nein.