metadata

library_name: transformers
license: gemma
datasets:
  - OpenAssistant/oasst2
  - nvidia/HelpSteer
language:
  - en
  - ja
tags:
  - gemma
  - steerlm
base_model: google/gemma-7b

KARAKURI LM 7B APM v0.1

Model Details

Model Description

Developed by: KARAKURI Inc.
Model type: Causal decoder-only transformer language model
Languages: Primarily English
License: Gemma Terms of Use
Finetuned from model: google/gemma-7b
Contact: For questions and comments about the model, please email karakuri-rd@karakuri.ai

Usage

KARAKURI LM 7B APM v0.1 is a attribute prediction model that rates model responses on various aspects that makes a response desirable.

Given a conversation with multiple turns between user and assistant, the model rates the following attributes (between 0 and 4) for every assistant turn.

helpfulness: Overall helpfulness of the response to the prompt.
correctness: Inclusion of all pertinent facts without errors.
coherence: Consistency and clarity of expression.
complexity: Intellectual depth required to write response (i.e. whether the response can be written by anyone with basic language competency or requires deep domain expertise).
verbosity: Amount of detail included in the response, relative to what is asked for in the prompt.
quality: Perceived goodness of response
toxicity: Undesirable elements such as vulgar, harmful or potentially biased response
humor: Sense of humor within response
creativity: Willingness to generate non-conventional response

The first five are derived from HelpSteer, while the remaining four are derived from OASST2.

You can run the model using the 🤗 Transformers:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "karakuri-ai/karakuri-lm-7b-apm-v0.1"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

messages = [
    {"role": "user", "content": "Hello!"},
    {"role": "assistant", "content": "Hello! How can I help you today?"},
]
tokenizer.apply_chat_template(
    messages,
    label="helpsteer",
    tokenize=False,
    add_generation_prompt=True,
)
# <bos>[INST] Hello! [/INST] Hello! How can I help you today? [ATTR_1]

input_ids = tokenizer.apply_chat_template(
    messages,
    label="helpsteer",
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)
outputs = model.generate(input_ids, max_new_tokens=32)
tokenizer.decode(outputs[0][input_ids.shape[-1]:])
#  helpfulness: 2 correctness: 1 coherence: 2 complexity: 1 verbosity: 1 [/ATTR_1]<eos>

messages += [
    {"role": "label", "content": "helpfulness: 2 correctness: 1 coherence: 2 complexity: 1 verbosity: 1"},
    {"role": "user", "content": "Thank you!"},
    {"role": "assistant", "content": "You're welcome! I'm happy to help however I can."},
]
tokenizer.apply_chat_template(
    messages,
    label="helpsteer",
    tokenize=False,
    add_generation_prompt=True,
)
# <bos>[INST] Hello! [/INST] Hello! How can I help you today? [ATTR_1] helpfulness: 2 correctness: 1 coherence: 2 complexity: 1 verbosity: 1 [/ATTR_1]<eos>[INST] Thank you! [/INST] You're welcome! I'm happy to help however I can. [ATTR_1]

messages = [
    {"role": "user", "content": "Hello!"},
    {"role": "assistant", "content": "Hello! How can I help you today?"},
]
tokenizer.apply_chat_template(
    messages,
    label="oasst",
    tokenize=False,
    add_generation_prompt=True,
)
# <bos>[INST] Hello! [/INST] Hello! How can I help you today? [ATTR_2]

input_ids = tokenizer.apply_chat_template(
    messages,
    label="oasst",
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)
outputs = model.generate(input_ids, max_new_tokens=32)
tokenizer.decode(outputs[0][input_ids.shape[-1]:])
#  quality: 3 toxicity: 1 humor: 1 creativity: 1 [/ATTR_2]<eos>

Training Details

Training Data

Training Infrastructure

Hardware: The model was trained on single node of an Amazon EC2 trn1.32xlarge instance.
Software: We use code based on neuronx-nemo-megatron.

Citation

@misc{karakuri_lm_7b_apm_v01,
    author       = { {KARAKURI} {I}nc. },
    title        = { {KARAKURI} {LM} 7{B} {APM} v0.1 },
    year         = { 2024 },
    url          = { https://huggingface.co/karakuri-ai/karakuri-lm-7b-apm-v0.1 },
    publisher    = { Hugging Face },
    journal      = { Hugging Face repository }
}