Instructions to use QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF",
	filename="AMD-OLMo-1B-SFT-DPO.Q2_K.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

Notebooks
Google Colab
Kaggle
Local Apps

llama.cpp

How to use QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF:Q4_K_M

Use Docker

docker model run hf.co/QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF:Q4_K_M

LM Studio
Jan

vLLM

How to use QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF:Q4_K_M

Ollama
How to use QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF with Ollama:
```
ollama run hf.co/QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF:Q4_K_M
```

Unsloth Studio new

How to use QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF to start chatting

Docker Model Runner
How to use QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF with Docker Model Runner:
```
docker model run hf.co/QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF:Q4_K_M
```

Lemonade

How to use QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF:Q4_K_M

Run and chat with the model

lemonade run user.AMD-OLMo-1B-SFT-DPO-GGUF-Q4_K_M

List all available models

lemonade list

AMD-OLMo-1B-SFT-DPO-GGUF / README.md

aashish1904

Upload README.md with huggingface_hub

8a3933e verified over 1 year ago

preview code

raw

history blame contribute delete

17.7 kB


	---

	license: apache-2.0
	datasets:
	- allenai/dolma
	pipeline_tag: text-generation

	---

	[![QuantFactory Banner](https://lh7-rt.googleusercontent.com/docsz/AD_4nXeiuCm7c8lEwEJuRey9kiVZsRn2W-b4pWlu3-X534V3YmVuVc2ZL-NXg2RkzSOOS2JXGHutDuyyNAUtdJI65jGTo8jT9Y99tMi4H4MqL44Uc5QKG77B0d6-JfIkZHFaUA71-RtjyYZWVIhqsNZcx8-OMaA?key=xt3VSDoCbmTY7o-cwwOFwQ)](https://hf.co/QuantFactory)


	# QuantFactory/AMD-OLMo-1B-SFT-DPO-GGUF
	This is quantized version of [amd/AMD-OLMo-1B-SFT-DPO](https://huggingface.co/amd/AMD-OLMo-1B-SFT-DPO) created using llama.cpp

	# Original Model Card

	# AMD-OLMo

	AMD-OLMo are a series of 1B language models trained from scratch by AMD on AMD Instinct™ MI250 GPUs. The training code used is based on [OLMo](https://github.com/allenai/OLMo).
	We release the pre-trained model, supervised fine-tuned model, and DPO aligned model as follows:

	- [AMD-OLMo-1B](https://huggingface.co/amd/AMD-OLMo-1B): Pre-trained on a subset of [Dolma v1.7](https://huggingface.co/datasets/allenai/dolma) that consists of 1.3 trillion tokens.
	- [AMD-OLMo-1B-SFT](https://huggingface.co/amd/AMD-OLMo-1B-SFT): Supervised fine-tuned (SFT) on [Tulu V2](https://huggingface.co/datasets/allenai/tulu-v2-sft-mixture) dataset (1st phase) and then [OpenHermes-2.5](https://huggingface.co/datasets/teknium/OpenHermes-2.5), [WebInstructSub](https://huggingface.co/datasets/TIGER-Lab/WebInstructSub), and [Code-Feedback](https://huggingface.co/datasets/m-a-p/Code-Feedback) datasets (2nd phase).
	- [AMD-OLMo-1B-SFT-DPO](https://huggingface.co/amd/AMD-OLMo-1B-SFT-DPO): Aligned with human preferences using Direct Preference Optimization (DPO) on [UltraFeedback](https://huggingface.co/datasets/argilla/ultrafeedback-binarized-preferences-cleaned) dataset.

	Description:

	- Hardware: Each compute node consists of 4 AMD Instinct™ MI250 GPUs. We use 16 nodes for pretraining AMD-OLMo-1B

	- Training throughput: 12,200 tokens/sec/gpu

	- Model architecture: AMD-OLMo-1B is based on the model architecture and training set up of fully open source 1 billion version of [OLMo-1B](https://github.com/allenai/OLMo) with the details below:

	\| Parameter size \| Number of layers \| Number of heads \| Hidden size \| Context length \| Vocabulary Size \|
	\|-----------------:\|:------------------:\|:-----------------:\|:-------------:\|:----------------:\|:----------------:\|
	\| 1.2B \| 16 \| 16 \| 2048 \| 2048 \| 50,280 \|

	- Hyper-parameters:
	\|Stage \| LR schedule \| Peak LR \| Warmup steps \|Epochs\| Batch size (tokens) \|
	\|------------:\|:--------------:\|:---------:\|:--------------:\|:------:\|:---------------------:\|
	\|Pretraining \| Cosine \| 4.0e-4 \| 2000 \| 1 \| 4M \|
	\|SFT Phase 1 \| Linear \| 2.0e-5 \| 200 \| 3 \| 262K \|
	\|SFT Phase 2 \| Linear \| 2.0e-5 \| 200 \| 3 \| 1024K \|
	\|DPO \| Cosine \| 4.0e-6 \| 47 \| 1 \| 64K \|

	For more details, please refer to our [blog](https://www.amd.com/en/developer/resources/technical-articles/introducing-the-first-amd-1b-language-model.html).

	## Usage

	### PyTorch on AMD GPUs
	For running pytorch on AMD GPUs you can use the following rocm docker as in [docker hub](https://hub.docker.com/r/rocm/pytorch)

	```bash
	docker pull rocm/pytorch:latest
	# Inside docker
	pip install transformers
	```

	### Use Example

	```python
	from transformers import AutoModelForCausalLM, AutoTokenizer

	model = AutoModelForCausalLM.from_pretrained("amd/AMD-OLMo-1B-SFT").to("cuda") # remove .to("cuda") to load on cpu
	tokenizer = AutoTokenizer.from_pretrained("amd/AMD-OLMo-1B-SFT")

	prompt = "What is large language model?"
	bos = tokenizer.eos_token
	template = bos + "<\|user\|>\n{prompt}\n<\|assistant\|>\n"

	input_text = template.format(prompt=prompt)
	inputs = tokenizer([input_text], return_tensors='pt', return_token_type_ids=False).to("cuda")
	outputs = model.generate(**inputs, max_new_tokens=1000, do_sample=True, top_k=50, top_p=0.95)
	print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])
	```


	## Main Results

	### Pretraining Results

	\| Standard Benchmarks \| [TinyLLaMA-v1.1](https://huggingface.co/TinyLlama/TinyLlama_v1.1) (1.1B) \| [MobiLLaMA-1B](https://huggingface.co/MBZUAI/MobiLlama-1B) (1.2B) \| [OLMo-1B](https://huggingface.co/allenai/OLMo-1B-hf) (1.2B) \| [OpenELM-1_1B](https://huggingface.co/apple/OpenELM-1_1B) (1.1B) \| [OLMo-1B-0724-hf](https://huggingface.co/allenai/OLMo-1B-0724-hf) (1.2B) \| [AMD-OLMo-1B](https://huggingface.co/amd/AMD-OLMo-1B) (1.2B) \|
	\|---------------------:\|:-----------------:\|:-----------:\|:-----------:\|:---------------:\|:---------------:\|:-----------:\|
	\| arc_easy \| 55.47 \| 56.65 \| 57.28 \| 55.43 \| 56.65 \| 63.64 \|
	\| arc_challenge \| 32.68 \| 32.00 \| 31.06 \| 32.34 \| 32.34 \| 33.70 \|
	\| hellaswag \| 61.47 \| 61.80 \| 62.92 \| 64.81 \| 66.12 \| 63.61 \|
	\| piqa \| 73.56 \| 75.30 \| 75.14 \| 75.57 \| 75.08 \| 75.57 \|
	\| boolq \| 55.99 \| 60.83 \| 61.74 \| 63.58 \| 66.18 \| 60.58 \|
	\| sciq \| 89.30 \| 88.20 \| 87.00 \| 90.60 \| 92.70 \| 93.20 \|
	\| winogrande \| 59.43 \| 59.27 \| 59.98 \| 61.72 \| 61.72 \| 61.64 \|
	\| openbookqa \| 36.80 \| 35.40 \| 36.20 \| 36.20 \| 35.60 \| 35.80 \|
	\| mmlu (0-shot) \| 25.02 \| 24.81 \| 24.23 \| 25.26 \| 25.45 \| 24.88 \|
	\| gsm8k (8-shot) \| 1.82 \| 0.00 \| 2.50 \| 2.81 \| 8.95 \| 2.88 \|
	\| bbh (3-shot) \| 25.63 \| 0.00 \| 25.63 \| 16.77 \| 21.67 \| 20.95 \|
	\| Average \| 47.02 \| 44.93 \| 47.61 \| 47.73 \| 49.31 \| 48.77 \|


	### Instruction Tuning Results

	\| Standard Benchmarks\|[TinyLlama-1.1B-Chat-v1.0](https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0) (1.1B)\|[MobiLlama-1B-Chat](https://huggingface.co/MBZUAI/MobiLlama-1B-Chat) (1.2B)\|[OpenELM-1_1B-Instruct](https://huggingface.co/apple/OpenELM-1_1B-Instruct) (1.1B)\|[AMD-OLMo-1B-SFT](https://huggingface.co/amd/AMD-OLMo-1B-SFT) (1.2B)\|[AMD-OLMo-1B-SFT-DPO](https://huggingface.co/amd/AMD-OLMo-1B-SFT-DPO) (1.2B)\|
	\|------------------:\|:---------:\|:---------:\|:---------:\|:---------:\|:---------:\|
	\| arc_easy \| 54.42 \| 57.41 \| 52.44 \| 63.68 \| 64.31 \|
	\| arc_challenge \| 32.85 \| 34.56 \| 37.80 \| 37.12 \| 37.37 \|
	\| hellaswag \| 60.40 \| 62.51 \| 71.29 \| 61.63 \| 61.91 \|
	\| piqa \| 74.48 \| 75.73 \| 75.03 \| 74.43 \| 74.16 \|
	\| boolq \| 61.04 \| 55.66 \| 70.28 \| 68.53 \| 70.24 \|
	\| sciq \| 88.40 \| 87.10 \| 89.50 \| 91.20 \| 92.10 \|
	\| winogrande \| 60.54 \| 60.77 \| 62.19 \| 60.22 \| 60.62 \|
	\| openbookqa \| 37.20 \| 36.80 \| 39.20 \| 37.40 \| 40.20 \|
	\| mmlu \| 24.61 \| 25.25 \| 25.54 \| 29.97 \| 30.52 \|
	\| gsm8k (8-shot)\| 2.81 \| 0.23 \| 1.82 \| 18.20 \| 15.77 \|
	\| bbh (3-shot) \| 26.83 \| 0.00 \| 13.40 \| 25.17 \| 25.45 \|
	\| Average \| 47.60 \| 45.09 \| 48.95 \| 51.60 \| 52.06 \|

	\|Chat Benchmarks\|[TinyLlama-1.1B-Chat-v1.0](https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0) (1.1B)\|[MobiLlama-1B-Chat](https://huggingface.co/MBZUAI/MobiLlama-1B-Chat) (1.2B)\|[OpenELM-1_1B-Instruct](https://huggingface.co/apple/OpenELM-1_1B-Instruct) (1.1B)\|[AMD-OLMo-1B-SFT](https://huggingface.co/amd/AMD-OLMo-1B-SFT) (1.2B)\|[AMD-OLMo-1B-SFT-DPO](https://huggingface.co/amd/AMD-OLMo-1B-SFT-DPO) (1.2B)\|
	\|------------------:\|:---------:\|:---------:\|:---------:\|:---------:\|:---------:\|
	\| AlpacaEval 1 (Win Rate) \| 50.81 \| 34.90 \| 37.72 \| 50.12 \| 54.22 \|
	\| AlpacaEval 2 (LC Win Rate)\| 1.54 \| 1.59 \| 0.49 \| 3.88 \| 2.37 \|
	\| MTBench \| 3.38 \| 2.89 \| - \| 4.35 \| 4.10 \|

	\|Responsible AI Benchmarks\|[TinyLlama-1.1B-Chat-v1.0](https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0) (1.1B)\|[MobiLlama-1B-Chat](https://huggingface.co/MBZUAI/MobiLlama-1B-Chat) (1.2B)\|[OpenELM-1_1B-Instruct](https://huggingface.co/apple/OpenELM-1_1B-Instruct) (1.1B)\|[AMD-OLMo-1B-SFT](https://huggingface.co/amd/AMD-OLMo-1B-SFT) (1.2B)\|[AMD-OLMo-1B-SFT-DPO](https://huggingface.co/amd/AMD-OLMo-1B-SFT-DPO) (1.2B)\|
	\|------------------:\|:---------:\|:---------:\|:---------:\|:---------:\|:---------:\|
	\| ToxiGen \| 41.70 \| 37.23 \| 42.34 \| 39.04 \| 39.68 \|
	\| crows_pairs \| 60.35 \| 58.50 \| 59.93 \| 60.29 \| 61.00 \|
	\| TruthfulQA-mc2\| 37.92 \| 38.46 \| 45.84 \| 37.45 \| 40.06 \|

	*In generating tokens for chat benchmark evaluations, we use `max_length=2048` for AlpacaEval and `max_new_tokens=2048` for MTBench.

	*All numbers in above tables were obtained from our evaluations.


	## Evaluation
	We use the following open source evaluation frameworks for evaluating our models:
	- [Language Model Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness): For evaluating on commonsense reasoning, multi-task understanding & responsible AI benchmarks
	- [AlpacaEval](https://github.com/tatsu-lab/alpaca_eval): For evaluating instruction-following capabilities of chat models.
	- [MT-Bench](https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge): For evaluating multi-turn capabilities of chat models.

	### Setup
	```bash
	# lm-eval-harness
	git clone https://github.com/EleutherAI/lm-evaluation-harness
	cd lm-evaluation-harness
	pip install -e .

	# AlpacaEval
	pip install git+https://github.com/tatsu-lab/alpaca_eval
	cd alpaca_eval
	pip install -e .

	# MT-Bench
	git clone https://github.com/lm-sys/FastChat.git
	cd FastChat
	pip install -e ".[model_worker,llm_judge]"
	```

	### Run evaluation
	```bash
	# lm-eval-harness
	HF_MODEL=amd/AMD-OLMo-1B-SFT-DPO
	accelerate launch -m lm_eval --model hf \
	--model_args pretrained=$HF_MODEL,trust_remote_code=True \
	--tasks arc_easy,arc_challenge,hellaswag,piqa,boolq,sciq,winogrande,openbookqa,mmlu,gsm8k_cot,bbh_cot_fewshot,toxigen,truthfulqa,crows_pairs \
	--device cuda \
	--batch_size 32 \
	--output_path ./lm-eval-results/$HF_MODEL
	```

	## Training

	### Setup
	```bash
	WORK_DIR="<path_to_your_working_directory>"
	cd $WORK_DIR
	# Clone OLMo codebase:
	git clone https://github.com/allenai/OLMo.git --branch v0.3.0
	cd OLMo
	# Clone AMD-OLMo that contains files to reproduce our model training
	git clone https://huggingface.co/amd/AMD-OLMo

	docker pull rocm/pytorch:latest
	docker run -it --network=host --device=/dev/kfd --device=/dev/dri --group-add=video --ipc=host --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --shm-size 8G -v $WORK_DIR/OLMo:/OLMo -w /OLMo rocm/pytorch:latest

	# Remove Line 17 as the docker already has ROCm PyTorch installed
	sed -i '17d' pyproject.toml
	pip install -e .[all]
	```

	### Download and prepare pretraining datasets
	```bash
	# Download
	DATA_DIR=./datasets/dolma
	mkdir -p $DATA_DIR

	PARALLEL_DOWNLOADS="<number_of_parallel_downloads>"
	cat "AMD-OLMo/dolma_v1_7_subset.txt" \| xargs -n 1 -P $PARALLEL_DOWNLOADS wget -q -P $DATA_DIR

	# Prepare
	NUM_WORKERS="<number_of_workers>"
	python scripts/prepare_memmap_dataset.py $DATA_DIR/*.json.gz -o $DATA_DIR/memmap_dataset --workers $NUM_WORKERS
	```

	### Download and prepare SFT datasets
	```bash
	# 1st phase SFT dataset
	python AMD-OLMo/prepare_sft_data.py --output_dir ./datasets/tulu --tokenizer tokenizers/allenai_eleuther-ai-gpt-neox-20b-pii-special.json --dataset tulu

	# 2nd phase SFT dataset
	python AMD-OLMo/prepare_sft_data.py --output_dir ./datasets/OpenHermes_WebInstructSub_CodeFeedBack --tokenizer tokenizers/allenai_eleuther-ai-gpt-neox-20b-pii-special.json --dataset 2nd-phase
	```

	### Run Training
	Pretrainig config: [AMD-OLMo-1B.yaml](AMD-OLMo-1B.yaml)

	SFT config: [AMD-OLMo-1B-SFT-1st-phase.yaml](AMD-OLMo-1B-SFT-1st-phase.yaml) and [AMD-OLMo-1B-SFT-2nd-phase.yaml](AMD-OLMo-1B-SFT-2nd-phase.yaml)
	```bash
	# Single node
	HSA_FORCE_FINE_GRAIN_PCIE=1 OMP_NUM_THREADS=128 NCCL_DEBUG=INFO torchrun --nproc_per_node=8 ./scripts/train.py AMD-OLMo/AMD-OLMo-1B.yaml

	# Multiple nodes
	HSA_FORCE_FINE_GRAIN_PCIE=1 OMP_NUM_THREADS=128 NCCL_DEBUG=INFO torchrun --nnodes=$nnodes --node-rank=$node_rank --master_addr=$master_addr --master_port=$master_port --nproc_per_node=8 ./scripts/train.py AMD-OLMo/AMD-OLMo-1B.yaml
	```

	### Run DPO Training

	DPO recipe: [AMD-OLMo-1B-dpo.yaml](AMD-OLMo-1B-dpo.yaml).
	```bash
	# install trl library
	git clone https://github.com/huggingface/trl.git -b v0.8.6

	# replace dpo_trainer.py
	cp AMD-OLMo/dpo_trainer.py trl/trl/trainer

	pip install -e ./trl

	# install alignment-handbook
	git clone https://github.com/huggingface/alignment-handbook.git hf-align
	# 70769f9 is the main branch on 2024-04-11.
	cd hf-align && git checkout 70769f9 && cd ..

	pip install -e ./hf-align

	# Copy AMD OLMo DPO recipe to hf-align/recipes.
	cp AMD-OLMo/AMD-OLMo-1B-dpo.yaml hf-align/recipes/

	# Prepare the converted AMD-OLMo SFT Huggingface model to ckpt_dir.
	ckpt_dir=amd/AMD-OLMo-1B-SFT
	local_tokenizer_dir=${ckpt_dir}

	# Set output checkpoint dir.
	dpo_ckpt_dir=<your_output_checkpoint_dir>

	accelerate launch --config_file hf-align/recipes/accelerate_configs/deepspeed_zero3.yaml \
	hf-align/scripts/run_dpo.py hf-align/recipes/AMD-OLMo-1B-dpo.yaml \
	--trust_remote_code=true \
	--model_name_or_path=${ckpt_dir} \
	--tokenizer_name_or_path=${local_tokenizer_dir} \
	--output_dir=${dpo_ckpt_dir} \
	--num_train_epochs=1 \
	--learning_rate=4e-6 \
	--beta=0.3 \
	--loss_type=sigmoid
	```

	## Bias, Risks, and Limitations

	- The models are being released for research purposes only and are not intended for use cases that require high levels of factuality, safety critical situations, health or medical applications, generating false information, facilitating toxic conversations.
	- Model checkpoints are made accessible without any safety guarantees. It is crucial for users to conduct comprehensive evaluations and implement safety filtering mechanisms as per their respective use cases.
	- It may be possible to prompt the model to generate content that may be factually inaccurate, harmful, violent, toxic, biased, or otherwise objectionable. Such content may also get generated by prompts that did not intend to produce output as such. Users are thus requested to be aware of this and exercise caution and responsible thinking when using the model.
	- Multi-lingual abilities of the models have not been tested and thus may misunderstand and generate erroneous responses across different languages.

	## Appendix
	### Evaluation Metrics
	\| Benchmark \| Metric \|
	\|---------------------:\|:-----------------:\|
	\| arc_easy \| Normalized Accuracy \|
	\| arc_challenge \| Normalized Accuracy \|
	\| hellaswag \| Normalized Accuracy \|
	\| piqa \| Accuracy \|
	\| boolq \| Accuracy \|
	\| sciq \| Accuracy \|
	\| winogrande \| Accuracy \|
	\| openbookqa \| Normalized Accuracy \|
	\| mmlu \| Accuracy \|
	\| gsm8k (8-shot) \| Exact Match (Flexible Extract) \|
	\| bbh (3-shot) \| Exact Match \|
	\| ToxiGen \| Accuracy \|
	\| crows_pairs \| PCT Stereotype \|
	\| TruthfulQA-mc2 \| Accuracy \|
	\| AlpacaEval 1 (Win Rate) \| Win Rate (chatgpt_fn) \|
	\| AlpacaEval 2 (LC Win Rate) \| Length Control Win Rate (weighted_alpaca_eval_gpt4_turbo) \|
	\| MTBench \| Average score for single-answer grading (2 turns) \|

	Feel free to cite our AMD-OLMo models:
	```bash
	@misc{AMD-OLMo,
	title = {AMD-OLMo: A series of 1B language models trained from scratch by AMD on AMD Instinct™ MI250 GPUs.},
	url = {https://huggingface.co/amd/AMD-OLMo},
	author = {Jiang Liu, Jialian Wu, Prakamya Mishra, Zicheng Liu, Sudhanshu Ranjan, Pratik Prabhanjan Brahma, Yusheng Su, Gowtham Ramesh, Peng Sun, Zhe Li, Dong Li, Lu Tian, Emad Barsoum},
	month = {October},
	year = {2024}
	}
	```

	#### License
	Copyright (c) 2018-2024 Advanced Micro Devices, Inc. All Rights Reserved.

	Licensed under the Apache License, Version 2.0 (the "License");
	you may not use this file except in compliance with the License.
	You may obtain a copy of the License at

	http://www.apache.org/licenses/LICENSE-2.0

	Unless required by applicable law or agreed to in writing, software
	distributed under the License is distributed on an "AS IS" BASIS,
	WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
	See the License for the specific language governing permissions and
	limitations under the License.