See axolotl config

axolotl version: 0.3.0

base_model: mistralai/Mistral-7B-Instruct-v0.2
model_type: MistralForCausalLM
tokenizer_type: LlamaTokenizer
is_mistral_derived_model: true

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: out/train_alpaca.jsonl
    type:
      alpaca
dataset_prepared_path:
val_set_size: 0.05
output_dir: ./mistral_fine_out

sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true

wandb_project:
wandb_entity:
wandb_watch:
wandb_run_id:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 4
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.000005

train_on_inputs: false
group_by_length: false
bf16: true
fp16: false
tf32: false

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

warmup_steps: 10
eval_steps: 0.05
eval_table_size:
eval_table_max_new_tokens: 128
save_steps:
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  bos_token: "<s>"
  eos_token: "</s>"
  unk_token: "<unk>"
model_config:
  sliding_window: 4096

The fine tuning script used for launch was from https://github.com/totallylegitco/healthinsurance-llm w/ run_remote.sh and an INPUT_MODEL=mistral

TotallyLegitCo/fighthealthinsurance_model_v0.3

This model is a fine-tuned version of mistralai/Mistral-7B-Instruct-v0.2 on the syntehtic-appeal dataset. It achieves the following results on the evaluation set:

• Loss: 1.3954

Model description

Generate health insurance appeals. Early work.

Intended uses & limitations

Generate health insurance appeals. This is early work and may not be suitable for production.

Training and evaluation data

The syntehtic appeal dataset was used for training and evaluation. Given how the dataset was produced there is likely cross-contamination of the training and eval datasets so loss values are likely understated.

This model is intended to match the Mistral-7B-Instruct style with <s>[INST]Instructions[/INT] present (as well as system specific instructions within an extra <<SYS><</SYS>.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-06
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• num_epochs: 4

Training results

Training Loss	Epoch	Step	Validation Loss
2.0506	0.0	1	2.4510
0.8601	0.2	58	1.1493
0.8635	0.4	116	1.1356
0.869	0.61	174	1.1174
0.7764	0.81	232	1.1173
0.7803	1.01	290	1.1124
0.6902	1.2	348	1.1570
0.6774	1.4	406	1.1591
0.6859	1.6	464	1.1651
0.725	1.81	522	1.1677
0.6525	2.01	580	1.1686
0.5069	2.2	638	1.2688
0.4702	2.4	696	1.2767
0.4888	2.6	754	1.2852
0.5197	2.8	812	1.2881
0.4734	3.01	870	1.2851
0.3586	3.2	928	1.3856
0.3889	3.4	986	1.3929
0.3526	3.6	1044	1.3959
0.3832	3.8	1102	1.3954

Framework versions

• Transformers 4.37.0.dev0
• Pytorch 2.0.1
• Datasets 2.16.1
• Tokenizers 0.15.0