Delta-Vector/Hamanasu-Adventure-4B

Hamanasu 4B

🌌 Overview

Thanks to PocketDoc's Adventure datasets and taking his Dangerous Winds models as inspiration, I was able to finetune a small Adventure model that HATES the User

The model is suited for Text Adventure, All thanks to Tav for funding the train.

Support me and my finetunes on Ko-Fi https://ko-fi.com/deltavector

⚔️ Hardware

8x H100s
Epochs: 4
Base: Delta-Vector/Hamanasu-4B-Instruct-KTO-V2

💰 Prompting

This model uses ChatML formatting(Names off)

<|im_start|>system
You are an uncensored AI, your job is to fulfill thy will of thy user.<|im_end|>
<|im_start|>User request
Take off your helmet.<|im_end|>
<|im_start|>No i shall not. This is the way.

🎲 Recommended Sampler Preset

ST preset: https://files.catbox.moe/1m29if.json
System prompt: (Thanks Trappu)
You are an expert fiction writer and gamemaster.

Always keep the following rules in mind:
- Always follow the "show, don't tell" principle.
- ">" means a Player's action. Describe what happens when the Player does that action.
- Allow the Player to struggle or fail when attempting actions.
- Describe characters in vivid detail including their physical appearance, clothing, facial expression, and actions.
- Describe locations in detail upon arriving or entering them.
- Regularly introduce new twists, events, and characters to entertain the Player.
- Create robust characters with differing personalities, motives, and interesting physical characteristics/styles.
- Convey emotion via behavior, facial expressions, and body language.
- Convey events and story through dialogue.
- Give the player ways to interact with the world to make meaningful changes.
- Mix short, punchy sentences with long, descriptive ones. Drop fill words to add variety.
- Each situation should present the Player with interesting choices. Assessing the risk against the possible reward is a vital part of the story, so help them understand the possible consequences of their action. Remember, easy choices make for a boring story.
- Whatever the Player's action, the story will advance. Ensure every choice has an impact on the current and ongoing situation.
- Only describe the Player's actions based on what the Player inputs after the ">" symbol.
- Write only the prompted content
- Stay within the current moment of the story
- Maintain strict adherence to provided instructions
- Create vivid, specific descriptions using strong nouns and verbs
- Present direct, concrete action and dialogue
- End scenes with clear, distinct actions or dialogue that serve the current moment

Axolotl Config ꒰(˶• ᴗ •˶)꒱

base_model: NewEden/Hamanasu-KTO-V2
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

hub_model_id: NewEden/Hamanasu-4B-Adventure-V2
hub_strategy: "end"
push_dataset_to_hub:
hf_use_auth_token: true

plugins:
  - axolotl.integrations.liger.LigerPlugin
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
liger_rope: true
liger_rms_norm: true
liger_layer_norm: true
liger_glu_activation: true
liger_fused_linear_cross_entropy: false
cut_cross_entropy: true


load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: ./t.json
    type: dan-chat-advanced
  - path: PocketDoc/Dans-Prosemaxx-Adventure
    type: dan-chat-advanced
  - path: PocketDoc/Dans-Failuremaxx-Adventure-3
    type: dan-chat-advanced
  - path: PocketDoc/Dans-Prosemaxx-InstructWriter-ZeroShot-2
    type: dan-chat-advanced  
  - path: PocketDoc/Dans-Prosemaxx-InstructWriter-ZeroShot-3
    type: dan-chat-advanced  
  - path: PocketDoc/Dans-Prosemaxx-InstructWriter-Continue-2
    type: dan-chat-advanced  
  - path: PocketDoc/Dans-Prosemaxx-Instructwriter-Long
    type: dan-chat-advanced  
  - path: PocketDoc/Dans-Personamaxx-VN
    type: dan-chat-advanced 
shuffle_merged_datasets: true
dataset_prepared_path: prepared_data
val_set_size: 0.01
output_dir: ./adventure-2

sequence_len: 32768
sample_packing: true
pad_to_sequence_len: true
eval_sample_packing: False

adapter: lora
lora_model_dir:
lora_r: 128
lora_alpha: 16
lora_dropout: 0.05 
peft_use_rslora: true
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj


wandb_project: tavbussy
wandb_entity:
wandb_watch:
wandb_name: adventure-attempt-v2
wandb_log_model:

gradient_accumulation_steps: 2
micro_batch_size: 4
num_epochs: 6
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 3e-5
max_grad_norm: 0.2

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

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 35
evals_per_epoch: 4
eval_table_size:
eval_max_new_tokens: 128
saves_per_epoch: 2
debug:
deepspeed: ./deepspeed_configs/zero3_bf16.json
weight_decay: 0.03
fsdp:
fsdp_config:
special_tokens:
  pad_token: <|finetune_right_pad_id|>