Hamanasu 15B Instruct
## π 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!)
### π Quantizations
| Type | Link |
|:---:|:---:|
| `GGUF` | https://huggingface.co/Delta-Vector/Hamanasu-15B-Instruct-gguf/ |
| `EXL2` | https://huggingface.co/Delta-Vector/Hamanasu-15B-Instruct-exl2 |
### βοΈ Hardware
- 4x RTX 3090 GPUs
- Epochs: 4
- Base: `Hamanasu-15B-R2-PT`
- Amount of Tokens: 1+ Billion
## π° 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
```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 κ°(ΛΆβ’ α΄ β’ΛΆ)κ±
```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:
```
## β‘ Credits
---
Made by
Delta-Vector