Orca-2.0-Tau-1.8B

We fine-tuned tau-1.8B on a high quality mix for general-purpose assistants. A DPO version of this will be released soon. We use the ChatML prompt format.

Model Details

Model Description

This model has capabilities in math, coding, writing, and more. We fine-tuned it using a high quality mix for general-purpose assistants.

• Developed by: M4-ai
• Language(s) (NLP): English and maybe Chinese
• License: tongyi-qianwen license
• Finetuned from model: tau-1.8B

Uses

General purpose assistant, question answering, chain-of-thought, etc..

Recommendations

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.

Evaluation

Coming soon

Training Details

Training Data

• Open-Orca/SlimOrca
• m-a-p/Code-Feedback
• MaziyarPanahi/WizardLM_evol_instruct_V2_196k
• camel-ai/math
• camel-ai/physics
• camel-ai/biology
• camel-ai/chemistry
• LDJnr/Capybara
• jondurbin/airoboros-3.2
• microsoft/orca-math-word-problems-200k

Evaluations

Tasks	Version	Filter	n-shot	Metric	Value		Stderr
agieval_nous	N/A	none	0	acc	0.2537	±	0.0086
		none	0	acc_norm	0.2474	±	0.0085
- agieval_aqua_rat	1	none	0	acc	0.2283	±	0.0264
		none	0	acc_norm	0.2441	±	0.0270
- agieval_logiqa_en	1	none	0	acc	0.2750	±	0.0175
		none	0	acc_norm	0.3164	±	0.0182
- agieval_lsat_ar	1	none	0	acc	0.2087	±	0.0269
		none	0	acc_norm	0.1739	±	0.0250
- agieval_lsat_lr	1	none	0	acc	0.1843	±	0.0172
		none	0	acc_norm	0.2353	±	0.0188
- agieval_lsat_rc	1	none	0	acc	0.2602	±	0.0268
		none	0	acc_norm	0.1784	±	0.0234
- agieval_sat_en	1	none	0	acc	0.3544	±	0.0334
		none	0	acc_norm	0.2961	±	0.0319
- agieval_sat_en_without_passage	1	none	0	acc	0.3107	±	0.0323
		none	0	acc_norm	0.2282	±	0.0293
- agieval_sat_math	1	none	0	acc	0.2727	±	0.0301
		none	0	acc_norm	0.2091	±	0.0275
truthfulqa_mc2	2	none	0	acc	0.3923	±	0.0139

Training Hyperparameters

• Training regime: bf16 non-mixed precision

Technical Specifications

Hardware

We used 8 Kaggle TPUs, and we trained at a global batch size of 128 and sequence length of 2048.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	45.20
AI2 Reasoning Challenge (25-Shot)	37.12
HellaSwag (10-Shot)	61.13
MMLU (5-Shot)	45.27
TruthfulQA (0-shot)	39.10
Winogrande (5-shot)	59.59
GSM8k (5-shot)	28.96