Mawdistical/Mawdistic-Purrvance-LoRa

See axolotl config

axolotl version: 0.7.0

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

model_config:
  trust_remote_code: true
  tokenizer:
    pad_token: "</s>"
    padding_side: "right"

datasets:
  - path: data/data.jsonl
    type: chat_template
    field_messages: conversations
    message_field_role: role
    message_field_content: content


load_in_4bit: true
adapter: qlora
lora_r: 64
lora_alpha: 32
lora_dropout: 0.1
lora_target_modules:
  - q_proj
  - k_proj
  - v_proj
  - o_proj
  - gate_proj
  - up_proj
  - down_proj

wandb_project: mawzilla
wandb_name: Purring-stats

val_set_size: 0.01
evals_per_epoch: 2
eval_sample_packing: false
eval_max_new_tokens: 128

lora_modules_to_save:
  - embed_tokens
  - lm_head


bf16: true
flash_attention: true
flash_attn_implementation: "flash_attention_2"
gradient_checkpointing: true
deepspeed: deepspeed_configs/zero3.json

gradient_accumulation_steps: 4
micro_batch_size: 8
num_epochs: 1.5
optimizer: paged_adamw_32bit
lr_scheduler: cosine
learning_rate: 1e-5
warmup_ratio: 0.03

max_seq_length: 8192
pad_to_sequence_len: false
sample_packing: false

output_dir: ./output
save_steps: 100
logging_steps: 10
save_safetensors: true

special_tokens:
  pad_token: "<|finetune_right_pad_id|>"
  eos_token: "<|eot_id|>"

output

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

• Loss: 2.1094

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: 1e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• optimizer: Use paged_adamw_32bit 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: 13
• num_epochs: 1.5

Training results

Training Loss	Epoch	Step	Validation Loss
No log	0.0035	1	2.8304
2.1716	0.5017	145	2.1812
2.1088	1.0035	290	2.1094

Framework versions

• PEFT 0.14.0
• Transformers 4.48.3
• Pytorch 2.5.1+cu121
• Datasets 3.2.0
• Tokenizers 0.21.1