NovusResearch/Thestral7b-v0.1

Thestral v0.1

Thestral is Mistral Fine-tune. The model is a QLoRA version of mistralai/Mistral-7B-v0.1 on the Open-Orca/SlimOrca.

This model is finetuned using 1xH100 using axolotl.

See axolotl config

axolotl version: 0.4.0

base_model: mistralai/Mistral-7B-v0.1
model_type: MistralForCausalLM
tokenizer_type: LlamaTokenizer

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: Open-Orca/SlimOrca
    type: sharegpt
dataset_prepared_path: last_run_prepared
val_set_size: 0.1
output_dir: ./qlora-out_2

adapter: qlora
lora_model_dir:

sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true

lora_r: 128
lora_alpha: 32
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj

wandb_project: slim_orca
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 1
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

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

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

loss_watchdog_threshold: 5.0
loss_watchdog_patience: 3

warmup_steps: 10
evals_per_epoch: 4
eval_table_size:
eval_max_new_tokens: 128
saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  bos_token: "<s>"
  eos_token: "</s>"
  unk_token: "<unk>"

GPT-4All Benchmark Set

Tasks	Version	Filter	n-shot	Metric	Value		Stderr
winogrande	1	none	None	acc	0.7498	±	0.0122
piqa	1	none	None	acc	0.8172	±	0.0090
		none	None	acc_norm	0.8286	±	0.0088
openbookqa	1	none	None	acc	0.3380	±	0.0212
		none	None	acc_norm	0.4420	±	0.0222
hellaswag	1	none	None	acc	0.6254	±	0.0048
		none	None	acc_norm	0.8061	±	0.0039
boolq	2	none	None	acc	0.8740	±	0.0058
arc_easy	1	none	None	acc	0.8199	±	0.0079
		none	None	acc_norm	0.7891	±	0.0084
arc_challenge	1	none	None	acc	0.5145	±	0.0146
		none	None	acc_norm	0.5461	±	0.0145
Average: 71.93

🤖 Additional information about training

This model is fine-tuned for 1.0 epoch.

Loss graph

Thanks to axolotl for making the repository we used to make this model.