NewEden/Francois-CKPTs

See axolotl config

axolotl version: 0.8.0

base_model: PocketDoc/Dans-PersonalityEngine-V1.1.0-12b
## 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

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: Personamaxx-VN.json
    type: dan-chat-advanced
  - path: NewEden/LIMARP-Complexity
    type: dan-chat-advanced
  - path: NewEden/PIPPA-Mega-Filtered
    type: dan-chat-advanced
  - path: NewEden/OpenCAI-ShareGPT
    type: dan-chat-advanced
  - path: NewEden/Creative_Writing-Complexity
    type: dan-chat-advanced
  - path: NewEden/Light-Novels-Roleplay-Logs-Books-Oh-My-duplicate-turns-removed
    type: dan-chat-advanced
  - path: prosemaxx-adventure-failuremaxx.json
    type: dan-chat-advanced
  - path: NewEden/Books-V2-ShareGPT
    type: dan-chat-advanced
  - path: NewEden/Deepseek-V3-RP-Filtered
    type: dan-chat-advanced
  - path: NewEden/BlueSky-10K-Complexity
    type: dan-chat-advanced
  - path: NewEden/Final-Alpindale-LNs-ShareGPT
    type: dan-chat-advanced
  - path: NewEden/DeepseekRP-Filtered
    type: dan-chat-advanced 
  - path: NewEden/RP-logs-V2-Experimental
    type: dan-chat-advanced 
  - path: anthracite-org/kalo_opus_misc_240827
    type: dan-chat-advanced 
  - path: anthracite-org/kalo_misc_part2
    type: dan-chat-advanced 
  - path: NewEden/vanilla-backrooms-claude-sharegpt
    type: dan-chat-advanced 
  - path: NewEden/Storium-Prefixed-Clean
    type: dan-chat-advanced 


## LOra so we dont fuck brains
adapter: lora
lora_model_dir:
lora_r: 128
lora_alpha: 16
lora_dropout: 0.05 
peft_use_rslora: true
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
shuffle_merged_datasets: true
dataset_prepared_path: prepared_data
output_dir: ./output/Francois-V2


## Ctx Length
sequence_len: 16384
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: false
#batch_flattening: true

#torch_compile: auto  # Optional[Union[Literal["auto"], bool]]
#torch_compile_backend:  # Optional[str]
## Wandb
wandb_project: Francois
wandb_entity:
wandb_watch:
wandb_name: v3
wandb_log_model:

## Hparams
gradient_accumulation_steps: 2
micro_batch_size: 2
num_epochs: 4
optimizer: paged_ademamix_8bit
lr_scheduler: cosine
learning_rate: 3e-5
max_grad_norm: 0.0001
weight_decay: 0.02
warmup_steps: 40

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

## Unsloth is broken, Use grad-ckpting.
gradient_checkpointing: true
early_stopping_patience:
#resume_from_checkpoint: /home/ubuntu/Mango/axolotl/outputs/checkpoint-1088
local_rank:
logging_steps: 1
xformers_attention: False
flash_attention: True
s2_attention:


## Evals
val_set_size: 0.0025
evals_per_epoch: 4
eval_table_size:
eval_max_new_tokens: 128
saves_per_epoch: 2
debug:
# Multi-GPU
deepspeed: ./deepspeed_configs/zero2.json
fsdp:
fsdp_config:
special_tokens:
  pad_token: <pad>

output/Francois-V2

This model is a fine-tuned version of PocketDoc/Dans-PersonalityEngine-V1.1.0-12b on the Personamaxx-VN.json, the NewEden/LIMARP-Complexity, the NewEden/PIPPA-Mega-Filtered, the NewEden/OpenCAI-ShareGPT, the NewEden/Creative_Writing-Complexity, the NewEden/Light-Novels-Roleplay-Logs-Books-Oh-My-duplicate-turns-removed, the prosemaxx-adventure-failuremaxx.json, the NewEden/Books-V2-ShareGPT, the NewEden/Deepseek-V3-RP-Filtered, the NewEden/BlueSky-10K-Complexity, the NewEden/Final-Alpindale-LNs-ShareGPT, the NewEden/DeepseekRP-Filtered, the NewEden/RP-logs-V2-Experimental, the anthracite-org/kalo_opus_misc_240827, the anthracite-org/kalo_misc_part2, the NewEden/vanilla-backrooms-claude-sharegpt and the NewEden/Storium-Prefixed-Clean datasets. It achieves the following results on the evaluation set:

• Loss: 2.1779

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: 3e-05
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• total_eval_batch_size: 16
• optimizer: Use paged_ademamix_8bit and the args are: No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 40
• num_epochs: 4.0

Training results

Training Loss	Epoch	Step	Validation Loss
1.731	0.0023	1	2.4143
1.451	0.2506	109	2.3014
1.4026	0.5011	218	2.2824
1.6573	0.7517	327	2.2581
1.587	1.0023	436	2.2424
1.2928	1.2529	545	2.2229
1.4023	1.5034	654	2.2034
1.6312	1.7540	763	2.1959
1.3044	2.0046	872	2.1909
1.4984	2.2552	981	2.1876
1.3767	2.5057	1090	2.1840
1.3972	2.7563	1199	2.1812
1.3663	3.0069	1308	2.1792
1.4958	3.2575	1417	2.1785
1.4214	3.5080	1526	2.1784
1.4001	3.7586	1635	2.1779

Framework versions

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