dolly-v2-7b Model Card

Summary

Databricks’ dolly-v2-7b, a istructio-followig large laguage model traied o the Databricks machie learig platform that is licesed for commercial use. Based o pythia-6.9b, Dolly is traied o ~15k istructio/respose fie tuig records databricks-dolly-15k geerated by Databricks employees i capability domais from the IstructGPT paper, icludig braistormig, classificatio, closed QA, geeratio, iformatio extractio, ope QA ad summarizatio. dolly-v2-7b is ot a state-of-the-art model, but does exhibit surprisigly high quality istructio followig behavior ot characteristic of the foudatio model o which it is based.

Dolly v2 is also available i these other models sizes:

• dolly-v2-12b, a 12 billio parameter based o pythia-12b
• dolly-v2-3b, a 2.8 billio parameter based o pythia-2.8b

Please refer to the dolly GitHub repo for tips o ruig iferece for various GPU cofiguratios.

Ower: Databricks, Ic.

Model Overview

dolly-v2-7b is a 6.9 billio parameter causal laguage model created by Databricks that is derived from EleutherAI’s Pythia-6.9b ad fie-tued o a ~15K record istructio corpus geerated by Databricks employees ad released uder a permissive licese (CC-BY-SA)

Usage

To use the model with the trasformers library o a machie with GPUs, first make sure you have the trasformers ad accelerate libraries istalled. I a Databricks otebook you could ru:

%pip istall accelerate>=0.12.0 trasformers[torch]==4.25.1

The istructio followig pipelie ca be loaded usig the pipelie fuctio as show below. This loads a custom IstructioTextGeeratioPipelie foud i the model repo here, which is why trust_remote_code=True is required. Icludig torch_dtype=torch.bfloat16 is geerally recommeded if this type is supported i order to reduce memory usage. It does ot appear to impact output quality. It is also fie to remove it if there is sufficiet memory.

示例代码

from modelscope.pipelies import pipelie
from modelscope.utils.costat import Tasks

if __ame__ == '__mai__':
    model = "AI-ModelScope/dolly-v2-7b"
    pipe = pipelie(Tasks.text_geeratio, model=model, model_revisio='v1.0.1', device='cuda:0',max_legth=40)
    istructio = "Explai to me the differece betwee uclear fissio ad fusio."
    output = pipe(istructio)
    prit(output)

Kow Limitatios

Performace Limitatios

dolly-v2-7b is ot a state-of-the-art geerative laguage model ad, though quatitative bechmarkig is ogoig, is ot desiged to perform competitively with more moder model architectures or models subject to larger pretraiig corpuses.

The Dolly model family is uder active developmet, ad so ay list of shortcomigs is ulikely to be exhaustive, but we iclude kow limitatios ad misfires here as a meas to documet ad share our prelimiary fidigs with the commuity.
I particular, dolly-v2-7b struggles with: sytactically complex prompts, programmig problems, mathematical operatios, factual errors, dates ad times, ope-eded questio aswerig, halluciatio, eumeratig lists of specific legth, stylistic mimicry, havig a sese of humor, etc. Moreover, we fid that dolly-v2-7b does ot have some capabilities, such as well-formatted letter writig, preset i the origial model.

Dataset Limitatios

Like all laguage models, dolly-v2-7b reflects the cotet ad limitatios of its traiig corpuses.

• The Pile: GPT-J’s pre-traiig corpus cotais cotet mostly collected from the public iteret, ad like most web-scale datasets, it cotais cotet may users would fid objectioable. As such, the model is likely to reflect these shortcomigs, potetially overtly i the case it is explicitly asked to produce objectioable cotet, ad sometimes subtly, as i the case of biased or harmful implicit associatios.

• databricks-dolly-15k: The traiig data o which dolly-v2-7b is istructio tued represets atural laguage istructios geerated by Databricks employees durig a period spaig March ad April 2023 ad icludes passages from Wikipedia as refereces passages for istructio categories like closed QA ad summarizatio. To our kowledge it does ot cotai obsceity, itellectual property or persoally idetifyig iformatio about o-public figures, but it may cotai typos ad factual errors. The dataset may also reflect biases foud i Wikipedia. Fially, the dataset likely reflects the iterests ad sematic choices of Databricks employees, a demographic which is ot represetative of the global populatio at large.

Databricks is committed to ogoig research ad developmet efforts to develop helpful, hoest ad harmless AI techologies that maximize the potetial of all idividuals ad orgaizatios.

Bechmark Metrics

Below you'll fid various models bechmark performace o the EleutherAI LLM Evaluatio Haress; model results are sorted by geometric mea to produce a itelligible orderig. As outlied above, these results demostrate that dolly-v2-7b is ot state of the art, ad i fact uderperforms dolly-v1-6b i some evaluatio bechmarks. We believe this owes to the compositio ad size of the uderlyig fie tuig datasets, but a robust statemet as to the sources of these variatios requires further study.

Happy Hackig!

dolly-v2-7b Model Card Summary Databricks’ dolly-v2-7b, an instruction-following large language mode