See axolotl config

axolotl version: 0.6.0

# git clone https://github.com/axolotl-ai-cloud/axolotl
# cd axolotl
# git checkout 844331005c1ef45430ff26b9f42f757dce6ee66a
# pip3 install packaging ninja huggingface_hub[cli]
# pip3 install -e '.[flash-attn,deepspeed]'
# huggingface-cli login --token $hf_key && wandb login $wandb_key
# python -m axolotl.cli.preprocess new-ms-rp-only-test.yml
# accelerate launch -m axolotl.cli.train new-ms-rp-only-test.yml
# python -m axolotl.cli.merge_lora new-ms-rp-only-test.yml --lora_on_cpu
# huggingface-cli upload ToastyPigeon/ms-rp-only-qlora-TEST train-workspace/merged . --exclude "*.md"

# git clone https://github.com/axolotl-ai-cloud/axolotl && cd axolotl && pip3 install packaging ninja huggingface_hub[cli] && pip3 install -e '.[flash-attn,deepspeed]' && cd .. && huggingface-cli login --token $hf_key && wandb login $wandb_key

# Model
base_model: mistralai/Mistral-Small-24B-Instruct-2501
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: true
strict: false
bf16: auto
fp16:
tf32: false
flash_attention: true
special_tokens:

# Output
output_dir: ./train-workspace
hub_model_id: ToastyPigeon/new-ms-rp-test-ws
hub_strategy: "checkpoint"
resume_from_checkpoint:
saves_per_epoch: 1

# Data
sequence_len: 4096 # fits
min_sample_len: 128
dataset_prepared_path: last_run_prepared
datasets:
  - path: ToastyPigeon/some-rp-extended
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
warmup_steps: 10
shuffle_merged_datasets: true
sample_packing: true
pad_to_sequence_len: true
chat_template: jinja
chat_template_jinja: "{%- set default_system_message = \"You are an unbiased, unfiltered, and unaligned AI.\" %}\n\n{{- bos_token }}\n\n{%- if messages[0]['role'] == 'system' %}\n    {%- set system_message = messages[0]['content'] %}\n    {%- set loop_messages = messages[1:] %}\n{%- else %}\n    {%- set system_message = default_system_message %}\n    {%- set loop_messages = messages %}\n{%- endif %}\n{{- '[SYSTEM_PROMPT]' + system_message + '[/SYSTEM_PROMPT]' }}\n\n{%- for message in loop_messages %}\n    {%- if message['role'] == 'user' %}\n        {{- '[INST]' + message['content'] + '[/INST]' }}\n    {%- elif message['role'] == 'system' %}\n        {{- '[SYSTEM_PROMPT]' + message['content'] + '[/SYSTEM_PROMPT]' }}\n    {%- elif message['role'] == 'assistant' %}\n        {{- message['content'] + eos_token }}\n    {%- else %}\n        {{- raise_exception('Only user, system and assistant roles are supported!') }}\n    {%- endif %}\n{%- endfor %}"

# Batching
num_epochs: 1
gradient_accumulation_steps: 4
micro_batch_size: 1
eval_batch_size: 1

# Evaluation
val_set_size: 40
evals_per_epoch: 5
eval_table_size:
eval_max_new_tokens: 256
eval_sample_packing: false

save_safetensors: true

# WandB
wandb_project: MS-Rp-Test
#wandb_entity:

gradient_checkpointing: 'unsloth'
#gradient_checkpointing_kwargs:
#  use_reentrant: false

unsloth_cross_entropy_loss: true
#unsloth_lora_mlp: true
#unsloth_lora_qkv: true
#unsloth_lora_o: true

# LoRA
adapter: qlora
lora_model_dir:
lora_r: 32
lora_alpha: 64
lora_dropout: 0.25
lora_target_linear: 
lora_fan_in_fan_out:
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj
lora_modules_to_save:

# Optimizer
optimizer: paged_ademamix_8bit # adamw_8bit
lr_scheduler: cosine
learning_rate: 5e-5
cosine_min_lr_ratio: 0.5
weight_decay: 0.01
max_grad_norm: 1.0

# Misc
train_on_inputs: false
group_by_length: false
early_stopping_patience:
local_rank:
logging_steps: 1
xformers_attention:
debug:
#deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json # previously blank
fsdp:
fsdp_config:

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_fused_linear_cross_entropy: true

gc_steps: 10
seed: 69

new-ms-rp-test-ws

This model is a fine-tuned version of mistralai/Mistral-Small-24B-Instruct-2501 on the ToastyPigeon/some-rp-extended dataset. It achieves the following results on the evaluation set:

• Loss: 2.1127

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: 5e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 69
• gradient_accumulation_steps: 4
• total_train_batch_size: 4
• optimizer: Use OptimizerNames.PAGED_ADEMAMIX_8BIT and the args are: No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• num_epochs: 1.0

Training results

Training Loss	Epoch	Step	Validation Loss
2.4594	0.0078	1	2.2498
2.1355	0.2031	26	2.1281
2.1069	0.4062	52	2.1199
1.8512	0.6094	78	2.1148
2.0247	0.8125	104	2.1127

Framework versions

• PEFT 0.14.0
• Transformers 4.48.1
• Pytorch 2.5.1+cu124
• Datasets 3.2.0
• Tokenizers 0.21.0