Hamanasu 32B
## π Overview
After 25 hours, I present Hamanasu-QwQ-32B-V0.1 - One of the first QwQ Finetunes. Using data from the following:
- `NewEden/Orion-LIT`
- `NewEden/Orion-Asstr-Stories-16K`
- `Mielikki/Erebus-87k`
This model shows great promise for roleplaying and story-writing. All thanks to Ruka-Hamanasu for funding the train.
Disclaimer: The model is still in preview, Only completion training has been performed ontop.
### π Quantizations
| Type | Link |
|:---:|:---:|
| `GGUF` | https://huggingface.co/Delta-Vector/Hamanasu-32B-V1-QwQ-exl2 |
| `EXL2` | https://huggingface.co/Delta-Vector/Hamanasu-32B-V1-QwQ-gguf |
### βοΈ Hardware
- 4x H100s
- Epochs: 1
- Base: `QwQ`
- Amount of Tokens: 1+ Billion
## π° Prompting
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 κ°(ΛΆβ’ α΄ β’ΛΆ)κ±
```yaml
base_model: Qwen/QwQ-32B
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: Mielikki/Erebus-87k
type: completion
field: body
- path: NewEden/Orion-Completion-Asstr-Stories-16K
type: completion
field: content
- path: NewEden/Orion-Completion-LIT
type: completion
field: text
shuffle_merged_datasets: true
dataset_prepared_path: prepared_data
output_dir: ./qvq-cum
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: qwq
wandb_entity:
wandb_watch:
wandb_name: Pretrain-pt1-v2-frfr
wandb_log_model:
gradient_accumulation_steps: 2
micro_batch_size: 2
num_epochs: 1
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 1e-5
max_grad_norm: 0.001
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: 40
saves_per_epoch: 2
debug:
deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
weight_decay: 0.01
fsdp:
fsdp_config:
```
## β‘ Credits
---
Made by
Delta-Vector