Test model to try to give an existing model QwQ's thoughts. For this first version it is ontop of PocketDoc/Dans-SakuraKaze-V1.0.0-12b (an rp/adventure/co-writing model), which was trained ontop of PocketDoc/Dans-PersonalityEngine-V1.1.0-12b (a jack of all trades instruct model), which was trained ontop of mistralai/Mistral-Nemo-Base-2407.

The prompt formatting and usage should be the same as with QwQ; Use ChatML, and remove the thinking from previous turns. If thoughts arent being generated automatically, add <think>\n to the start of the assistant turn.

It should follow previous model turns formatting. On first turns of the conversation you may need to regen a few times, and maybe edit the model responses for the first few turns to get it to your liking.

You may want to disable inserting {{char}}: prefix for the character, and instead add something like Only speak as "{{char}}" in conversation with "{{user}}". Output your final response with a "{{char}}:" prefix. to the end of you system prompt.

See axolotl config

axolotl version: 0.8.0.dev0

yaml

mlflow_tracking_uri: http://127.0.0.1:7860
mlflow_experiment_name: MN-2407-DSK-QwQify-v0.1-12B-LoRA

# Hugging Face saving config
hub_model_id: BeaverAI/MN-2407-DSK-QwQify-v0.1-12B-LoRA-WS
hub_strategy: every_save

# Model checkpointing config
output_dir: ./Outputs/MN-2407-DSK-QwQify-v0.1-12B-LoRA
resume_from_checkpoint:
save_steps: 25
save_safetensors: true
save_total_limit: 3
save_only_model: false

# Model architecture config
base_model: PocketDoc/Dans-SakuraKaze-V1.0.0-12b
model_type: MistralForCausalLM
tokenizer_type: PreTrainedTokenizerFast

# Mixed precision training config
bf16: true
fp16: false
tf32: false

# Model loading config
load_in_8bit: false
load_in_4bit: false
strict: false

# Sequence config
sequence_len: 8192
min_sample_len: 256
sample_packing: true
eval_sample_packing: true
pad_to_sequence_len: true
train_on_inputs: false
group_by_length: false

# LoRA adapter config
adapter: lora
lora_model_dir:
lora_r: 128
lora_alpha: 128
lora_dropout: 0.125
peft_layers_to_transform:
peft_use_dora:
peft_use_rslora:
peft_layer_replication:
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj
lora_modules_to_save:

# Fix uninitialized tokens (such as <|start_header_id|> on the base L3 models)
fix_untrained_tokens:

# Dataset config
# https://github.com/xzuyn/axolotl/blob/came-plus-formatters/src/axolotl/prompt_strategies/customchatml-regex-last-only.py
datasets:
  - path: PJMixers-Dev/allura-org_gryphe-sonnet-3.5-charcards-names-added-qwq-all-aphrodite-Shuffled
    split: train
    type: customchatml-regex-last-only
  - path: PJMixers-Dev/anthracite-org_c2_logs_32k_llama3_qwen2_v1.3-qwq-all-aphrodite-Shuffled
    split: train
    type: customchatml-regex-last-only
  - path: PJMixers-Dev/grimulkan_aicg-logs-augmented-system-qwq-all-aphrodite-Shuffled
    split: train
    type: customchatml-regex-last-only
  - path: PJMixers-Dev/grimulkan_jannie-log-augmented-system-qwq-all-aphrodite-Shuffled
    split: train
    type: customchatml-regex-last-only
  - path: PJMixers-Dev/grimulkan_PIPPA-augmented-dedup-system-qwq-all-aphrodite-Shuffled
    split: train
    type: customchatml-regex-last-only
  - path: PJMixers-Dev/lemonilia_LimaRP-Only-NonSus-Simple-CustomShareGPT-qwq-all-aphrodite-Shuffled
    split: train
    type: customchatml-regex-last-only
  - path: PJMixers-Dev/MinervaAI_Aesir-Preview-Anon-qwq-all-aphrodite-Shuffled
    split: train
    type: customchatml-regex-last-only
  - path: PJMixers-Dev/NyxKrage_chub-logs-sharegpt-longest-CustomShareGPT-qwq-all-aphrodite-Shuffled
    split: train
    type: customchatml-regex-last-only
  - path: PJMixers-Dev/PocketDoc_Dans-Prosemaxx-Cowriter-XL-8192-shrunk-l3-qwq-all-aphrodite-Shuffled
    split: train
    type: customchatml-regex-last-only
  - path: PJMixers-Dev/PocketDoc_Dans-Personamaxx-Rainy-qwq-all-aphrodite-Shuffled
    split: train
    type: customchatml-regex-last-only
test_datasets:
  - path: PJMixers-Dev/allura-org_gryphe-sonnet-3.5-charcards-names-added-qwq-all-aphrodite-Shuffled
    split: test
    type: customchatml-regex-last-only
  - path: PJMixers-Dev/anthracite-org_c2_logs_32k_llama3_qwen2_v1.3-qwq-all-aphrodite-Shuffled
    split: test
    type: customchatml-regex-last-only
  - path: PJMixers-Dev/grimulkan_aicg-logs-augmented-system-qwq-all-aphrodite-Shuffled
    split: test
    type: customchatml-regex-last-only
  - path: PJMixers-Dev/grimulkan_jannie-log-augmented-system-qwq-all-aphrodite-Shuffled
    split: test
    type: customchatml-regex-last-only
  - path: PJMixers-Dev/grimulkan_PIPPA-augmented-dedup-system-qwq-all-aphrodite-Shuffled
    split: test
    type: customchatml-regex-last-only
  - path: PJMixers-Dev/lemonilia_LimaRP-Only-NonSus-Simple-CustomShareGPT-qwq-all-aphrodite-Shuffled
    split: test
    type: customchatml-regex-last-only
  - path: PJMixers-Dev/MinervaAI_Aesir-Preview-Anon-qwq-all-aphrodite-Shuffled
    split: test
    type: customchatml-regex-last-only
  - path: PJMixers-Dev/NyxKrage_chub-logs-sharegpt-longest-CustomShareGPT-qwq-all-aphrodite-Shuffled
    split: test
    type: customchatml-regex-last-only
  - path: PJMixers-Dev/PocketDoc_Dans-Prosemaxx-Cowriter-XL-8192-shrunk-l3-qwq-all-aphrodite-Shuffled
    split: test
    type: customchatml-regex-last-only
  - path: PJMixers-Dev/PocketDoc_Dans-Personamaxx-Rainy-qwq-all-aphrodite-Shuffled
    split: test
    type: customchatml-regex-last-only
val_set_size: 0
eval_strategy: steps
eval_steps: 25
dataset_prepared_path: ./00-Tokenized-Datasets/MN-2407-DSK-QwQify-v0.1-12B-customchatml-regex-last-only
shuffle_merged_datasets: true
dataset_processes:

# Training hyperparameters
num_epochs: 2
gradient_accumulation_steps: 1
micro_batch_size: 16  # x4 GPUs = 64
eval_batch_size: 16   # x4 GPUs = 64
warmup_steps: 0
optimizer: came_pytorch
optim_args:
optim_target_modules:
lr_scheduler: rex
learning_rate: 2e-5
cosine_min_lr_ratio:
loraplus_lr_ratio:
loraplus_lr_embedding:
weight_decay: 0.1
max_grad_norm: 1
logging_steps: 1

# Model optimization
gradient_checkpointing: unsloth
flash_attention: true
plugins:
  - axolotl.integrations.liger.LigerPlugin
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
cut_cross_entropy: true
liger_rope: true
liger_rms_norm: true
liger_layer_norm: true
liger_glu_activation: true
liger_cross_entropy: false
liger_fused_linear_cross_entropy: false
lora_mlp_kernel: false
lora_qkv_kernel: false
lora_o_kernel: false

# DeepSpeed
deepspeed: deepspeed_configs/zero3_bf16.json

# Garbage Collection
gc_steps: 1

# Debug config
debug: true
seed: 42

# Token config
special_tokens:
  bos_token: "<s>"
  eos_token: "<|im_end|>"
  pad_token: "<pad>"
tokens:

MN-2407-DSK-QwQify-v0.1-12B-LoRA-WS

This model is a fine-tuned version of PocketDoc/Dans-SakuraKaze-V1.0.0-12b on the PJMixers-Dev/allura-org_gryphe-sonnet-3.5-charcards-names-added-qwq-all-aphrodite-Shuffled, the PJMixers-Dev/anthracite-org_c2_logs_32k_llama3_qwen2_v1.3-qwq-all-aphrodite-Shuffled, the PJMixers-Dev/grimulkan_aicg-logs-augmented-system-qwq-all-aphrodite-Shuffled, the PJMixers-Dev/grimulkan_jannie-log-augmented-system-qwq-all-aphrodite-Shuffled, the PJMixers-Dev/grimulkan_PIPPA-augmented-dedup-system-qwq-all-aphrodite-Shuffled, the PJMixers-Dev/lemonilia_LimaRP-Only-NonSus-Simple-CustomShareGPT-qwq-all-aphrodite-Shuffled, the PJMixers-Dev/MinervaAI_Aesir-Preview-Anon-qwq-all-aphrodite-Shuffled, the PJMixers-Dev/NyxKrage_chub-logs-sharegpt-longest-CustomShareGPT-qwq-all-aphrodite-Shuffled, the PJMixers-Dev/PocketDoc_Dans-Prosemaxx-Cowriter-XL-8192-shrunk-l3-qwq-all-aphrodite-Shuffled and the PJMixers-Dev/PocketDoc_Dans-Personamaxx-Rainy-qwq-all-aphrodite-Shuffled datasets. It achieves the following results on the evaluation set:

Loss: 1.2770

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: 16
eval_batch_size: 16
seed: 42
distributed_type: multi-GPU
num_devices: 4
total_train_batch_size: 64
total_eval_batch_size: 64
optimizer: Use OptimizerNames.ADAMW_HF with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: cosine
num_epochs: 2.0

Training results

Training Loss	Epoch	Step	Validation Loss
2.134	0.0038	1	2.0025
1.6185	0.0951	25	1.5748
1.5187	0.1901	50	1.4871
1.4757	0.2852	75	1.4410
1.4008	0.3802	100	1.4100
1.4116	0.4753	125	1.3857
1.357	0.5703	150	1.3630
1.3435	0.6654	175	1.3478
1.3332	0.7605	200	1.3353
1.3042	0.8555	225	1.3308
1.2993	0.9506	250	1.3228
1.3105	1.0456	275	1.3154
1.2782	1.1407	300	1.3094
1.3063	1.2357	325	1.3070
1.3003	1.3308	350	1.3005
1.2937	1.4259	375	1.2952
1.283	1.5209	400	1.2922
1.2692	1.6160	425	1.2887
1.2639	1.7110	450	1.2855
1.2546	1.8061	475	1.2822
1.2711	1.9011	500	1.2787
1.2492	1.9962	525	1.2770

Framework versions

PEFT 0.14.0
Transformers 4.49.0
Pytorch 2.6.0+cu124
Datasets 3.2.0
Tokenizers 0.21.1