🌌 Overview

After multiple days of training, I'm proud to showcase my very own Phi-4 Finetune, Pretrained on almost a billion tokens worth of Books from

NewEden/Orion-LIT
NewEden/Orion-Asstr-Stories-16K
Mielikki/Erebus-87k

Then instruct tuned. This model shows great promise for roleplaying while keeping things fresh and new. (And most of, Unsloppy!)

📈 Quantizations

Type	Link
`GGUF`	https://huggingface.co/Delta-Vector/Hamanasu-15B-Instruct-gguf/
`EXL2`	https://huggingface.co/Delta-Vector/Hamanasu-15B-Instruct-exl2

⚔️ Hardware

4x RTX 3090 GPUs
Epochs: 4
Base: Hamanasu-15B-R2-PT
Amount of Tokens: 1+ Billion

💰 Prompting

A known quirk of the model is overly verbose responses when generation length is uncapped, Please cap your maximum output tokens to 100~ tokens above what you prefer

This model uses ChatML formatting

python

<|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.

Axolotl Config ꒰(˶• ᴗ •˶)꒱

yaml

base_model: NewEden_Phi-PT-merged-LIT
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

plugins:
  - axolotl.integrations.liger.LigerPlugin
liger_rope: true
liger_rms_norm: true
liger_swiglu: true
liger_fused_linear_cross_entropy: true


load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: PocketDoc/Dans-MemoryCore-CoreCurriculum-Small
    type: dan-chat-advanced
  - path: Nitral-AI/ARES-ShareGPT
    type: dan-chat-advanced
  - path: Gryphe/Sonnet3.5-SlimOrcaDedupCleaned-20k
    type: dan-chat-advanced
  - path: NewEden/Claude-Instruct-2.7K
    type: dan-chat-advanced
  - path: NewEden/Claude-Instruct-5K
    type: dan-chat-advanced

shuffle_merged_datasets: true
dataset_prepared_path: prepared_data
val_set_size: 0.0
output_dir: ./phi4-inst-out-r2

sequence_len: 16384
sample_packing: true
pad_to_sequence_len: true

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

lora_modules_to_save:
 - embed_tokens
 - lm_head


wandb_project: mag-phi
wandb_entity:
wandb_watch:
wandb_name: inst-attempt-02
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 4
optimizer: paged_ademamix_8bit
lr_scheduler: cosine
learning_rate: 0.000025

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

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

warmup_steps: 15
evals_per_epoch: 4
eval_table_size:
eval_max_new_tokens: 128
saves_per_epoch: 2
debug:
deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16_cpuoffload_params.json
weight_decay: 0.01
fsdp:
fsdp_config: