L3.3 GeneticLemonade Unleashed V3 70B

🚀 GENETIC LEMONADE UNLEASHED v3

An experimental model designed for character - driven RP / ERP, offering long and narrative - rich responses.

📚 Documentation

🔍 Overview

• This is an experimental release. It is a SFT + DPO QLora finetune of zerofata/GeneticLemonade - Unleashed.
• It is a creative model aimed at excelling in character - driven RP / ERP. It has not been tested or trained on adventure stories or large - scale creative writing.
• The model is designed to provide longer, narrative - heavy responses where characters are portrayed accurately and proactively.

💻 SillyTavern Settings

• These settings are not the 'best' but a stable baseline. Note that this model supports higher temperatures than usually recommended for other L3 models.

Recommended Samplers

Instruct

Use Llama - 3 - Instruct - Names, but you need to uncheck "System same as user".

📦 Quantizations

GGUF

• iMatrix (Mradermacher)

EXL2

• 4bpw
• 4.5bpw
• 6bpw

🔧 Technical Details

Training Process

• The model first underwent SFT with a small synthetic dataset of 2.9 million tokens, approximately 750 conversations. It mainly consisted of RP data with small amounts of random instruct / assistant data and creative writing.
• Then, it went through DPO training using about 1100 chosen examples from the SFT dataset that were of exceptional quality or showed verifiable instruction following. Rejected samples were generated using another Llama 3.3 finetune known for poor instruction following.

Axolotl Configs

Neither config is optimized for cost / performance efficiency, and results may vary.

SFT 1*H200

# ====================
# MODEL CONFIGURATION
# ====================
base_model: zerofata/L3.3-GeneticLemonade-Unleashed-70B
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
special_tokens:
  pad_token: "<|finetune_right_pad_id|>"
chat_template: llama3

# ====================
# DATASET CONFIGURATION
# ====================
datasets:
  - path: ./dataset.jsonl
    type: chat_template
    split: train
    chat_template_strategy: tokenizer
    field_messages: messages
    message_property_mappings:
      role: role
      content: content
    roles:
      user: ["user"]
      assistant: ["assistant"]
      system: ["system"]

test_datasets:
  - path: ./validate_dataset.jsonl
    type: chat_template
    split: train
    chat_template_strategy: tokenizer
    field_messages: messages
    message_property_mappings:
      role: role
      content: content
    roles:
      user: ["user"]
      assistant: ["assistant"]
      system: ["system"]

dataset_prepared_path:
train_on_inputs: false  # Only train on assistant responses

# ====================
# QLORA CONFIGURATION
# ====================
adapter: qlora
load_in_4bit: true
lora_r: 64
lora_alpha: 128
lora_dropout: 0.1
lora_target_linear: true
# lora_modules_to_save:  # Uncomment only if you added NEW tokens

# ====================
# TRAINING PARAMETERS
# ====================
num_epochs: 2
micro_batch_size: 4
gradient_accumulation_steps: 2
learning_rate: 1.5e-5
optimizer: paged_adamw_8bit
lr_scheduler: rex
warmup_ratio: 0.05
weight_decay: 0.01
max_grad_norm: 1.0

# ====================
# SEQUENCE & PACKING
# ====================
sequence_len: 8192
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true

# ====================
# HARDWARE OPTIMIZATIONS
# ====================
bf16: auto
flash_attention: true
gradient_checkpointing: true

# ====================
# EVALUATION & CHECKPOINTING
# ====================
evaluation_strategy: steps
eval_steps: 5
save_strategy: steps
save_steps: 5
save_total_limit: 5  # Keep best + last few checkpoints
load_best_model_at_end: true
metric_for_best_model: eval_loss
greater_is_better: false
early_stopping_patience: 5

# ====================
# LOGGING & OUTPUT
# ====================
output_dir: ./output_model
logging_steps: 2
save_safetensors: true

# ====================
# WANDB TRACKING
# ====================
wandb_project: project_name
# wandb_entity: your_entity
# wandb_name: your_run_name

DPO 2*H200

# ====================
# MODEL CONFIGURATION
# ====================
base_model: ApocalypseParty/unleashed-fulldata30
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
special_tokens: {}
chat_template: tokenizer_default

# ====================
# RL/DPO CONFIGURATION
# ====================
rl: dpo
rl_beta: 0.07

# ====================
# DATASET CONFIGURATION
# ====================
datasets:
  - path: ./dpo_cleaned-v3_deduplicated.jsonl
    type: chat_template.default
    field_messages: conversation
    field_chosen: chosen
    field_rejected: rejected
    message_property_mappings:
      role: role
      content: content
    roles:
      system: ["system"]
      user: ["user"]
      assistant: ["assistant"]
dataset_prepared_path:
train_on_inputs: false  # Only train on assistant responses

# ====================
# QLORA CONFIGURATION
# ====================
adapter: qlora
load_in_4bit: true
lora_r: 32
lora_alpha: 64
lora_dropout: 0.05
lora_target_linear: true
# lora_modules_to_save:  # Uncomment only if you added NEW tokens

# ====================
# TRAINING PARAMETERS
# ====================
num_epochs: 1
micro_batch_size: 4
gradient_accumulation_steps: 2
learning_rate: 2e-6
optimizer: adamw_8bit
lr_scheduler: cosine
warmup_steps: 5
weight_decay: 0.01
max_grad_norm: 1.0

# ====================
# SEQUENCE CONFIGURATION
# ====================
sequence_len: 4096
pad_to_sequence_len: true

# ====================
# HARDWARE OPTIMIZATIONS
# ====================
bf16: auto
tf32: false
flash_attention: true
gradient_checkpointing: offload
deepspeed: deepspeed_configs/zero1.json

# ====================
# CHECKPOINTING
# ====================
save_steps: 10
save_total_limit: 10
load_best_model_at_end: true
metric_for_best_model: eval_loss
greater_is_better: false

# ====================
# LOG

📄 License

The model is under the llama3 license.