pszemraj/Mistral-7B-sarcasm-scrolls-v2

See axolotl config

axolotl version: 0.4.1

base_model: mistralai/Mistral-7B-v0.3
model_type: MistralForCausalLM
tokenizer_type: LlamaTokenizer

strict: false

# dataset
datasets:
    - path: BEE-spoke-data/sarcasm-scrolls
      type: completion # format from earlier
      field: text
val_set_size: 200

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true
train_on_inputs: false
group_by_length: false

# WANDB
wandb_project: sarcasm-scrolls
wandb_entity: pszemraj
wandb_watch: gradients
wandb_name: Mistral-7B-v0.3-sarcasm-scrolls-v2a
hub_model_id: pszemraj/Mistral-7B-v0.3-sarcasm-scrolls-v2
hub_strategy: every_save

gradient_accumulation_steps: 32
micro_batch_size: 1
num_epochs: 2
optimizer: adamw_torch_fused # paged_adamw_32bit
lr_scheduler: cosine
learning_rate: 2e-5

load_in_8bit: false
load_in_4bit: false
bf16: true
tf32: true

torch_compile: true 
torch_compile_backend: inductor # Optional[str]
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
early_stopping_patience:
logging_steps: 3
xformers_attention:
flash_attention: true

warmup_steps: 20
# hyperparams for freq of evals, saving, etc
evals_per_epoch: 4
saves_per_epoch: 4
save_safetensors: true
save_total_limit: 1 # Checkpoints saved at a time
output_dir: ./output-axolotl/output-model-chaz
resume_from_checkpoint:


deepspeed:
weight_decay: 0.06

special_tokens:

Mistral-7B-v0.3-sarcasm-scrolls-v2

Model description

This model is a fine-tuned version of mistralai/Mistral-7B-v0.3 on the BEE-spoke-data/sarcasm-scrolls dataset. It achieves the following results on the evaluation set:

• Loss: 2.3333

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• gradient_accumulation_steps: 32
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 20
• num_epochs: 2

Training results

Training Loss	Epoch	Step	Validation Loss
No log	0.0075	1	2.3935
2.3672	0.2548	34	2.3638
2.3751	0.5096	68	2.3499
2.308	0.7644	102	2.3238
2.2672	1.0035	136	2.3027
1.702	1.2583	170	2.3449
1.7456	1.5131	204	2.3370
1.7004	1.7679	238	2.3333

Framework versions

• Transformers 4.41.1
• Pytorch 2.3.1+cu118
• Datasets 2.19.1
• Tokenizers 0.19.1