gemma-7b-openhermes

gemma-7b-openhermes is a variant of the Gemma 7B language model, which has been further fine-tuned on the OpenHermes-2.5 preference dataset using QLoRA.

• google/gemma-7b-it
• mlabonne/chatml-OpenHermes2.5-dpo-binarized-alpha

Usage

Chat Template

The instruction-tuned models use a chat template that must be adhered to for conversational use. The easiest way to apply it is using the tokenizer's built-in chat template, as shown in the following snippet.

Let's load the model and apply the chat template to a conversation. In this example, we'll start with a single user interaction:

from transformers import AutoTokenizer, AutoModelForCausalLM
import transformers
import torch

model_id = "abideen/gemma-7b-openhermes"
dtype = torch.bfloat16

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="cuda",
    torch_dtype=dtype,
)

chat = [{ "role": "user", "content": "What is a Language Model?" }]
prompt = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)

After the prompt is ready, generation can be performed like this:

inputs = tokenizer.encode(prompt, add_special_tokens=True, return_tensors="pt")
outputs = model.generate(input_ids=inputs.to(model.device), max_new_tokens=250)
print(tokenizer.decode(outputs[0]))

Inputs and outputs

• Input: Text string, such as a question, a prompt, or a document to be summarized.
• Output: Generated English-language text in response to the input, such as an answer to a question, or a summary of a document.

🏆 Evaluation results

Nous Benchmark

Agieval

Task	Version	Metric	Value		StdErr
agieval_aqua_rat	0	acc	24.80	_	2.72
agieval_aqua_rat	0	acc_norm	24.80	_	2.72
agieval_logiqa_en	0	acc	20.89	_	1.59
agieval_logiqa_en	0	acc_norm	23.35	_	1.66
agieval_lsat_ar	0	acc	21.74	_	2.73
agieval_lsat_ar	0	acc_norm	20.43	_	2.66
agieval_lsat_lr	0	acc	15.49	_	1.60
agieval_lsat_lr	0	acc_norm	20.59	_	1.79
agieval_lsat_rc	0	acc	17.10	_	2.30
agieval_lsat_rc	0	acc_norm	17.84	_	2.34
agieval_sat_en	0	acc	29.61	_	3.19
agieval_sat_en	0	acc_norm	29.61	_	3.19
agieval_sat_en_without_passage	0	acc	26.21	_	3.07
agieval_sat_en_without_passage	0	acc_norm	24.76	_	3.01
agieval_sat_math	0	acc	22.73	_	2.83
agieval_sat_math	0	acc_norm	22.73	_	2.83
Average: 22.29

GPT4ALL

Task	Version	Metric	Value		StdErr
arc_challenge	0	acc	20.14	_	1.17
arc_challenge	0	acc_norm	22.87	_	1.23
arc_easy	0	acc	32.37	_	0.96
arc_easy	0	acc_norm	31.61	_	0.95
boolq	1	acc	45.78	_	0.87
hellaswag	0	acc	32.03	_	0.47
hellaswag	0	acc_norm	35.18	_	0.48
openbookqa	0	acc	17.8	_	1.71
openbookqa	0	acc_norm	29.8	_	2.05
piqa	0	acc	54.46	_	1.16
piqa	0	acc_norm	54.57	_	1.16
winogrande	0	acc	48.30	_	1.40
Average: 32.00

TruthfulQA

Task	Version	Metric	Value	Std Err
truthfulqa_mc	1	mc1	30.11	1.61
truthfulqa_mc	1	mc2	47.69	1.61
Average: 38.90

Openllm Benchmark

Task	Version	Metric	Value		Stderr
arc_challenge	0	acc	48.12	±	1.46
		acc_norm	51.27	±	1.46
hellaswag	0	acc	55.4	±	0.49
		acc_norm	71.92	±	0.42
gsm8k	0	acc	29.87	±	1.2
winogrande	0	acc	68.19	±	1.3
mmlu	0	acc	53.62	±	0.6

Average: 73.5%

TruthfulQA

Task	Version	Metric	Value		Stderr
truthfulqa_mc	1	mc1	30.23	±	1.60
		mc2	47.17	±	1.63

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-07
• train_batch_size: 1
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 8
• 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: 100
• training_steps: 1000

📝 Axolotl Configuration

base_model: google/gemma-7b-it
model_type: GemmaForCausalLM
tokenizer_type: GemmaTokenizer
trust_remote_code: true

load_in_8bit: false
load_in_4bit: true
strict: false

rl: dpo
chat_template: chatml
datasets:
  - path: mlabonne/chatml-OpenHermes2.5-dpo-binarized-alpha
    split: train
    type: chatml.intel
dataset_prepared_path:
val_set_size: 0.01
output_dir: ./out

adapter: qlora
lora_model_dir:

sequence_len: 1800
sample_packing: false
pad_to_sequence_len: false

lora_r: 16
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
lora_target_modules:

wandb_project: gemma
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 8
micro_batch_size: 1
num_epochs: 1
optimizer: paged_adamw_32bit
lr_scheduler: cosine
learning_rate: 5e-7

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

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

warmup_steps: 100
evals_per_epoch: 1
eval_table_size:
eval_table_max_new_tokens: 128
save_steps: 1000
max_steps: 1000
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:

Framework versions

• Transformers 4.39.0.dev0
• Pytorch 2.1.2+cu118
• Datasets 2.17.0
• Tokenizers 0.15.0
• axolotl: 0.4.0