Hamanasu 15B R2 PT
## 🌌 Overview
This is the 2nd pretrain of Phi-4 Continued from the Orginal Asstr-Erebus Pretrain. This pretrain used 500 million tokens from
- `NewEden/Orion-LIT`
This model has *not* been instruct tuned, Ablities to converse may be reduced from the original model, If you would like to roleplay, Please use the Instruct version.
### ⚔️ Hardware
- 4x RTX 3090 GPUs
- Epochs: 1
- Base: `Hamanasu-15B-R1-PT`
- Amount of Tokens: 500 Million
## Axolotl Config ꒰(˶• ᴗ •˶)꒱
```yaml
base_model: Hamanasu-15B-R2-PT
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
#hub_model_id: NewEden/Phi4-pretrain
#hub_strategy: "all_checkpoints"
#push_dataset_to_hub:
#hf_use_auth_token: true
plugins:
- axolotl.integrations.liger.LigerPlugin
liger_rope: true
liger_rms_norm: true
liger_swiglu: true
liger_fused_linear_cross_entropy: true
#plugins:
# - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
#cut_cross_entropy: true
load_in_8bit: false
load_in_4bit: false
strict: false
datasets:
- path: NewEden/Orion-LIT
type: completion
field: text
shuffle_merged_datasets: true
dataset_prepared_path: prepared_data
val_set_size: 0.0
output_dir: ./phi4-ptv2-out-r1
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: comp-v2-attempt-01
wandb_log_model:
gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 1
optimizer: paged_ademamix_8bit
lr_scheduler: cosine
learning_rate: 0.00002
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: 4
debug:
deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16_cpuoffload_params.json
weight_decay: 0.01
fsdp:
fsdp_config:
```
## ⚡ Credits
---
Made by
Delta-Vector