> For the complete documentation index, see [llms.txt](https://brindha.gitbook.io/mylearning/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://brindha.gitbook.io/mylearning/tools/orca.md).

# Orca

#### Microsoft Research Orca: Teaching Small Models to Reason

**What is Orca?**

Orca is a research project and model family from Microsoft Research focused on a central question: can small language models learn to reason as effectively as much larger ones if they are trained on the right kind of data? The Orca series demonstrated that by training smaller models on rich, detailed explanation traces generated by GPT-4 rather than simple input-output pairs, it is possible to produce models that match or exceed the performance of models five to ten times their size on reasoning benchmarks.

***

**Core Research Insight**

The fundamental insight behind Orca is that conventional instruction-tuning approaches teach models to produce outputs that look like those of larger models, but without the underlying reasoning process. When smaller models are fine-tuned only on final answers, they learn surface-level style rather than deep capabilities. Orca addressed this by training on explanation traces — step-by-step reasoning chains, chain-of-thought solutions, and detailed justifications generated by GPT-4 — rather than just final responses. This richer signal allows smaller models to internalize reasoning strategies rather than merely mimic output formats.

***

**Model Versions**

**Orca 1** — Released in 2023, fine-tuned from LLaMA 13B using complex explanation traces from GPT-4. Outperformed models like Vicuna-13B on complex tasks by a significant margin.

**Orca 2** — Released in November 2023, fine-tuned from LLaMA 2 in 7B and 13B sizes. Orca 2 introduced an additional innovation: training models to select the most appropriate reasoning strategy for a given task rather than always using the same approach. Available strategies included step-by-step reasoning, direct answering, recall-then-generate, and extraction-based responses. Orca 2 7B and 13B matched or surpassed LLaMA 2 Chat at 13B and 70B scales, as well as WizardLM at 13B and 70B, across a broad range of benchmarks covering language understanding, common-sense reasoning, multi-step reasoning, and mathematics.

**Orca 3 — AgentInstruct** — An evolution of the Orca methodology using agentic multi-step workflows to generate synthetic training data. When applied to Mistral 7B, AgentInstruct data produced a 40% improvement on AGIEval, 19% improvement on MMLU, 54% improvement on GSM8K, 38% improvement on BBH, and 45% improvement on AlpacaEval compared to the base Mistral Instruct model.

**Orca-Math** — A specialized 7B model trained on 200,000 high-quality synthetic math problems generated using AutoGen multi-agent pipelines. It outperforms LLaMA 2-70B, GPT-3.5, Gemini Pro, WizardMath-70B, and MetaMath-70B on the GSM8K math benchmark despite being a fraction of their size.

***

**License and Access**

Orca models are released for research purposes under non-commercial licenses, consistent with the LLaMA and LLaMA 2 base model restrictions they build upon.

***

**Strengths**

* Pioneering research on explanation-trace training as a path to stronger small model reasoning
* Orca 2's strategy selection — teaching models when to reason versus when to answer directly — is a sophisticated alignment contribution
* Exceptional benchmark performance relative to model size
* Orca-Math demonstrates the power of domain specialization with synthetic data
* Influential research methodology adopted by many subsequent projects

**Limitations**

* Non-commercial license
* Research models without production deployment support
* Based on older LLaMA generations
* No multimodal capabilities
* Primarily English