metadata

license: mit
datasets:
  - Open-Orca/SlimOrca-Dedup
  - migtissera/Synthia-v1.3
  - LDJnr/Verified-Camel
  - LDJnr/Pure-Dove
  - LDJnr/Capybara
  - meta-math/MetaMathQA
  - Intel/orca_dpo_pairs
  - argilla/ultrafeedback-binarized-preferences-cleaned

Phi-2 Orange

A two-step finetune of Phi-2, with a bit of zest.

There is an updated model at rhysjones/phi-2-orange-v2 which has higher evals, if you wish to test.

Training details

A first finetune using a collection of broad training data:

And then a DPO finetune using:

Run within Ollama

If you're using Ollama, you can download and run using:

ollama run rhysjones/phi-2-orange

Prompt Format

Phi-2 Orange uses ChatML as the prompt format, with or without the system instruction.

To prompt with a system instruction (use whatever system prompt you like):

<|im_start|>system
You are a helpful assistant for Python which outputs in Markdown format.<|im_end|>
<|im_start|>user
Write a function to calculate the Fibonacci sequence<|im_end|>
<|im_start|>assistant

You can also omit the system prompt if you wish:

<|im_start|>user
Why is the sky blue?<|im_end|>
<|im_start|>assistant

Evaluations

Evaluations done using mlabonne's usefull Colab notebook llm-autoeval. Also check out the alternative leaderboard at Yet_Another_LLM_Leaderboard

Model	AGIEval	GPT4All	TruthfulQA	Bigbench	Average
phi-2-orange	33.37	71.33	49.87	37.3	47.97
phi-2-dpo	30.39	71.68	50.75	34.9	46.93
dolphin-2_6-phi-2	33.12	69.85	47.39	37.2	46.89
phi-2	27.98	70.8	44.43	35.21	44.61