See axolotl config

axolotl version: 0.8.0

## model
base_model: hardlyworking/Broth-12B
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
## upload
hub_model_id: hardlyworking/Noodles-12B
hub_strategy: "all_checkpoints"
push_dataset_to_hub:
hf_use_auth_token: true
## qlora COPE
load_in_8bit: false
load_in_4bit: false
strict: false

## data 
datasets:
  - path: PocketDoc/Dans-MemoryCore-CoreCurriculum-Small
    type: dan-chat-advanced
  - path: NewEden/Kalo-Opus-Instruct-22k-Refusal-Murdered
    type: dan-chat-advanced
  - path: Epiculous/Synthstruct-Gens-v1.1-Filtered-n-Cleaned
    type: dan-chat-advanced
  - path: Nitral-AI/Reasoning-1shot_ShareGPT
    type: dan-chat-advanced
  - path: Nitral-AI/GU_Instruct-ShareGPT
    type: dan-chat-advanced
  - path: Nitral-AI/Medical_Instruct-ShareGPT
    type: dan-chat-advanced
  - path: AquaV/Resistance-Sharegpt
    type: dan-chat-advanced
  - path: AquaV/US-Army-Survival-Sharegpt
    type: dan-chat-advanced
  - path: Gryphe/Sonnet3.5-SlimOrcaDedupCleaned
    type: dan-chat-advanced

shuffle_merged_datasets: true
dataset_prepared_path: dataset_prepared
val_set_size: 0.001
output_dir: outputs/out

## LIGER & CCE
plugins:
  - axolotl.integrations.liger.LigerPlugin
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
liger_rope: true
liger_rms_norm: true
liger_layer_norm: true
liger_glu_activation: true
liger_fused_linear_cross_entropy: false
cut_cross_entropy: false

## CTX settings
sequence_len: 8192
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true

## Lora 
adapter: lora
lora_model_dir:
lora_r: 128
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
peft_use_rslora: true
lora_modules_to_save:
  - embed_tokens
  - lm_head

## WandB
wandb_project: JoeyBoy
wandb_entity:
wandb_watch:
wandb_name: 
wandb_log_model:

## evals
evals_per_epoch: 8
eval_table_size:
eval_max_new_tokens: 128

## hoe params
gradient_accumulation_steps: 2
micro_batch_size: 4
num_epochs: 2
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 2e-5

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
s2_attention:

warmup_steps: 40
saves_per_epoch: 2
debug:
## for ademiamix 
deepspeed: ./deepspeed_configs/zero3_bf16.json
## for adamw
#deepspeed: ./deepspeed_configs/zero3_bf16.json
weight_decay: 0.01
fsdp:
fsdp_config:
special_tokens:
   pad_token: <pad>

Noodles-12B

This model is a fine-tuned version of hardlyworking/Broth-12B on the PocketDoc/Dans-MemoryCore-CoreCurriculum-Small, the NewEden/Kalo-Opus-Instruct-22k-Refusal-Murdered, the Epiculous/Synthstruct-Gens-v1.1-Filtered-n-Cleaned, the Nitral-AI/Reasoning-1shot_ShareGPT, the Nitral-AI/GU_Instruct-ShareGPT, the Nitral-AI/Medical_Instruct-ShareGPT, the AquaV/Resistance-Sharegpt, the AquaV/US-Army-Survival-Sharegpt and the Gryphe/Sonnet3.5-SlimOrcaDedupCleaned datasets. It achieves the following results on the evaluation set:

• Loss: 0.7255

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• total_eval_batch_size: 16
• optimizer: Use OptimizerNames.PAGED_ADAMW_8BIT with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 40
• num_epochs: 2.0

Training results

Training Loss	Epoch	Step	Validation Loss
1.0128	0.0013	1	1.2742
0.8925	0.1256	94	0.8576
0.8065	0.2512	188	0.8102
0.782	0.3768	282	0.7881
0.8135	0.5023	376	0.7755
0.7877	0.6279	470	0.7601
0.7955	0.7535	564	0.7516
0.758	0.8791	658	0.7444
0.7362	1.0040	752	0.7402
0.7053	1.1296	846	0.7354
0.6439	1.2552	940	0.7326
0.7445	1.3808	1034	0.7298
0.5843	1.5063	1128	0.7288
0.6571	1.6319	1222	0.7268
0.652	1.7575	1316	0.7257
0.6872	1.8831	1410	0.7255

Framework versions

• PEFT 0.15.1
• Transformers 4.51.3
• Pytorch 2.6.0+cu124
• Datasets 3.5.0
• Tokenizers 0.21.1