Llama-2-13B-Instruct-v0.2

<div align="center"> <img src="./assets/llama.png" width="150px"> </div>

Llama-2-13B-Instruct-v0.2

A pretrained generative language model with 13 billion parameters geared towards instruction-following capabilities.

Model Details

This model was built via parameter-efficient finetuning of the meta-llama/Llama-2-13b-hf base model on the first 20k rows in each of the jondurbin/airoboros-2.2.1, Open-Orca/SlimOrca, and garage-bAInd/Open-Platypus datasets.

• Developed by: Daniel Furman
• Model type: Causal language model (clm)
• Language(s) (NLP): English
• License: Apache 2.0
• Finetuned from model: meta-llama/Llama-2-13b-hf

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Note: The below values do not apply the prompt formatting used to finetune the model. An action item for future development is to run these evaluation benchmarks with the formatting applied, which should increase the scores.

Basic Usage

<details> <summary>Setup</summary>

!pip install -q -U transformers peft torch accelerate einops sentencepiece

import torch
from peft import PeftModel, PeftConfig
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
)

peft_model_id = "dfurman/Llama-2-13B-Instruct-v0.2"
config = PeftConfig.from_pretrained(peft_model_id)

tokenizer = AutoTokenizer.from_pretrained(
    peft_model_id,
    use_fast=True,
    trust_remote_code=True,
)

model = AutoModelForCausalLM.from_pretrained(
    config.base_model_name_or_path,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

model = PeftModel.from_pretrained(
    model, 
    peft_model_id
)

</details>

messages = [
    {"role": "user", "content": "Tell me a recipe for a mai tai."},
]

print("\n\n*** Prompt:")
input_ids = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    return_tensors="pt",
)
print(tokenizer.decode(input_ids[0]))

print("\n\n*** Generate:")
with torch.autocast("cuda", dtype=torch.bfloat16):
    output = model.generate(
        input_ids=input_ids.cuda(),
        max_new_tokens=1024,
        do_sample=True,
        temperature=0.7,
        return_dict_in_generate=True,
        eos_token_id=tokenizer.eos_token_id,
        pad_token_id=tokenizer.pad_token_id,
        repetition_penalty=1.2,
        no_repeat_ngram_size=5,
    )

response = tokenizer.decode(
    output["sequences"][0][len(input_ids[0]):], 
    skip_special_tokens=True
)
print(response)

<details> <summary>Outputs</summary>

Prompt:

"<s> [INST] Tell me a recipe for a mai tai. [/INST]"

Generation:

"""1 ounce orange curaçao
4 ounces dark rum (I recommend using Myers's)
2 ounces light rum (Myers's is also good here)
3 ounces freshly squeezed lime juice
6 to 8 ice cubes
Shake all ingredients except the ice in a cocktail shaker until chilled, about 5 seconds. Add the ice and shake again until frosty, about another 5 seconds. Strain into an old-fashioned or double rocks glass filled with crushed ice. Garnish with an orange slice and a cherry speared on a skewer."""

</details>

Speeds, Sizes, Times

Training

It took ~8 hours to train 3 epochs on 1x A100 (40 GB SXM).

Prompt Format

This model was finetuned with the following format:

tokenizer.chat_template = "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST] ' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token + ' ' }}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}"

This format is available as a chat template via the apply_chat_template() method. Here's an illustrative example:

messages = [
    {"role": "user", "content": "Tell me a recipe for a mai tai."},
    {"role": "assistant", "content": "1 oz light rum\n½ oz dark rum\n¼ oz orange curaçao\n2 oz pineapple juice\n¾ oz lime juice\nDash of orgeat syrup (optional)\nSplash of grenadine (for garnish, optional)\nLime wheel and cherry garnishes (optional)\n\nShake all ingredients except the splash of grenadine in a cocktail shaker over ice. Strain into an old-fashioned glass filled with fresh ice cubes. Gently pour the splash of grenadine down the side of the glass so that it sinks to the bottom. Add garnishes as desired."},
    {"role": "user", "content": "How can I make it more upscale and luxurious?"},
]

print("\n\n*** Prompt:")
input_ids = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    return_tensors="pt",
)
print(tokenizer.decode(input_ids[0]))

<details> <summary>Output</summary>

"<s> [INST] Tell me a recipe for a mai tai. [/INST] 1 ounce orange curaçao\n4 ounces dark rum (...) orange slice and a cherry speared on a skewer.</s>  [INST] How can I make the mai tai more upscale and luxurious? [/INST] "

</details>

Training Hyperparameters

We use the SFTTrainer from trl to fine-tune LLMs on instruction-following datasets.

See here for the finetuning code, which contains an exhaustive view of the hyperparameters employed.

The following TrainingArguments config was used:

• output_dir = "./results"
• num_train_epochs = 2
• auto_find_batch_size = True
• gradient_accumulation_steps = 2
• optim = "paged_adamw_32bit"
• save_strategy = "epoch"
• learning_rate = 3e-4
• lr_scheduler_type = "cosine"
• warmup_ratio = 0.03
• logging_strategy = "steps"
• logging_steps = 25
• evaluation_strategy = "no"
• bf16 = True

The following bitsandbytes quantization config was used:

• quant_method: bitsandbytes
• load_in_8bit: False
• load_in_4bit: True
• llm_int8_threshold: 6.0
• llm_int8_skip_modules: None
• llm_int8_enable_fp32_cpu_offload: False
• llm_int8_has_fp16_weight: False
• bnb_4bit_quant_type: nf4
• bnb_4bit_use_double_quant: False
• bnb_4bit_compute_dtype: bfloat16

Model Card Contact

dryanfurman at gmail

Llama-2 Research Citation

@misc{touvron2023llama,
      title={Llama 2: Open Foundation and Fine-Tuned Chat Models}, 
      author={Hugo Touvron and Louis Martin and Kevin Stone and Peter Albert and Amjad Almahairi and Yasmine Babaei and Nikolay Bashlykov and Soumya Batra and Prajjwal Bhargava and Shruti Bhosale and Dan Bikel and Lukas Blecher and Cristian Canton Ferrer and Moya Chen and Guillem Cucurull and David Esiobu and Jude Fernandes and Jeremy Fu and Wenyin Fu and Brian Fuller and Cynthia Gao and Vedanuj Goswami and Naman Goyal and Anthony Hartshorn and Saghar Hosseini and Rui Hou and Hakan Inan and Marcin Kardas and Viktor Kerkez and Madian Khabsa and Isabel Kloumann and Artem Korenev and Punit Singh Koura and Marie-Anne Lachaux and Thibaut Lavril and Jenya Lee and Diana Liskovich and Yinghai Lu and Yuning Mao and Xavier Martinet and Todor Mihaylov and Pushkar Mishra and Igor Molybog and Yixin Nie and Andrew Poulton and Jeremy Reizenstein and Rashi Rungta and Kalyan Saladi and Alan Schelten and Ruan Silva and Eric Michael Smith and Ranjan Subramanian and Xiaoqing Ellen Tan and Binh Tang and Ross Taylor and Adina Williams and Jian Xiang Kuan and Puxin Xu and Zheng Yan and Iliyan Zarov and Yuchen Zhang and Angela Fan and Melanie Kambadur and Sharan Narang and Aurelien Rodriguez and Robert Stojnic and Sergey Edunov and Thomas Scialom},
      year={2023},
      eprint={2307.09288},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Framework versions

• PEFT 0.6.3.dev0

Open LLM Leaderboard Evaluation Results

Detailed results can be found here