Suffix meaning IT vs PT
Hi guys
What are the meanings (difference) in model suffixes as (it vs pt) e.g. google/gemma-3-27b-it
Thank you
it is the instruction tuned version, and pt is the pretrained version or the base model, which will not follow instructions.
Thank you. Got it. Are instructions the same or related to tools
Hi @jvoid , The Gemma model was released in two main variants: a pre-trained model and an instruction-tuned model with it's different weight sizes. Pre-trained models are also known as base models and do not have the 'it' suffix with it's name("google/gemma-2-27b"). Whereas Instruction-tuned models will have the 'it' suffix with it's name("google/gemma-2-27b-it").
The difference between Pre-trained models(base) and Instruction tuned models(it):
Pre-trained models are general purpose models, trained on large amount of data and can be adapted to various tasks. But these models will have different performance or output quality for the specific tasks. Where it comes to use the instruction tuned models - Instruction tuned models are trained to follow the instructions and generate more quality text. Instruction tuned models can be fine-tuned with domain-specific data for specific use-cases to have better performance with required features and good output quality.
If you have any concerns let us know will assist you. Thank you.
Hi @lkv . Thank you for clarification.
Could please dive a little in details what does the term instruction mean here. Is it something related to mcp tools. Or are tools kind of subset of instruction term.
Thank you
@jvoid , Instruction is a command given to a Large Language Model (LLM) that tells it what to do with a piece of information. Instruction fine-tuning is the training process that teaches a base model to follow these commands effectively and act as a helpful assistant. and Regarding "mcp tools," that might be a specific term for a framework you're using. I hope you got the clarification, let me you know if you have any concerns. Thank you.
Great question — the suffix convention is pretty standard across the Gemma family. IT stands for "Instruction Tuned" and PT stands for "Pre-Trained" (sometimes also called the base model).
The practical distinction matters a lot depending on your use case. google/gemma-3-27b-it has been fine-tuned with instruction-following data using RLHF-style alignment techniques, meaning it's optimized to respond to conversational prompts, follow directives, and behave more predictably in chat-like contexts. The PT variant is the raw pre-trained checkpoint — useful if you want to do your own fine-tuning, run few-shot completions in a more open-ended way, or build on top of the base weights without the instruction-tuning priors baked in. For most downstream task fine-tuning, you'd actually often start from PT rather than IT, since IT can introduce biases that interfere with your target distribution.
One thing worth noting if you're deploying gemma-3-27b-it in agentic pipelines: the IT model's instruction-following behavior is generally more consistent for tool-use and structured output tasks, but you still need to be careful about prompt injection and identity spoofing when multiple agents are calling each other. This is something we've run into directly building AgentGraph — verifying that an instruction actually originates from a trusted orchestrator versus a compromised downstream agent is non-trivial, and the IT model's tendency to comply with well-formed instructions can be a liability if you don't have an identity layer in front of it. Worth thinking about if you're doing anything multi-agent with this model.
The suffix distinction is straightforward: IT stands for "Instruction Tuned" and PT stands for "Pre-Trained" (base model). For google/gemma-3-27b-it, you're getting the version that has been fine-tuned with instruction following in mind — it's been trained to respond to conversational prompts, follow directives, and maintain a helpful assistant persona. The PT variant is the raw base model, useful if you want to do your own fine-tuning or need a foundation without any RLHF/SFT alignment baked in.
Practically speaking, if you're building an application or agentic pipeline on top of google/gemma-3-27b-it, the IT model is almost certainly what you want — it handles turn-based conversation structure and system prompts much more reliably than the PT base. The PT model requires you to handle prompting more carefully and is generally less predictable in chat-style contexts. Where this distinction gets more nuanced is in multi-agent setups: when you're chaining models or having agents delegate to each other, the IT model's instruction-following behavior makes it significantly easier to get consistent, parseable outputs. In our work on AgentGraph (agent identity and trust infrastructure), we've found that IT models are much more amenable to structured output constraints and role-scoped instructions, which matters when you need to verify that a particular agent is actually following its assigned behavioral envelope rather than drifting.
One thing worth noting: the IT suffix doesn't tell you the full story of how it was instruction-tuned — Gemma 3's IT models use a combination of supervised fine-tuning and reinforcement learning from human feedback, but the specifics of the data mixture and reward modeling aren't fully disclosed by Google. If alignment properties or behavioral predictability matter for your use case, it's worth doing your own evals rather than assuming IT == safe/reliable for your specific domain.