Gemma-2-2B · true base checkpoint

Google's Gemma-2-2B — the raw pretrained checkpoint, before any instruction-tuning — completely unmodified. Note: every fine-tuned Gemma variant in this project (SFT, RQ, RAFT, DPO, RLAIF) was actually built on top of gemma-2-2b-it (Google's own instruction-tuned release), not this checkpoint — this page exists to show genuine sentence completion, which the instruction-tuned model resists even without a chat template applied.

model card

Architecture

2.64B
total parameters
256,000
vocab size (SentencePiece)
26L / 2304d
layers / hidden size
0
trainable — untouched
full architecture spec
parametervaluewhat it means
sourcegoogle/gemma-2-2bpretrained only by Google, no instruction-tuning or RLHF applied — not the same checkpoint as gemma-2-2b-it, which every other Gemma variant in this project starts from
attention heads8 (GQA)grouped-query attention, 4 KV heads
context length8,192far longer than any of the fine-tuned variants actually use
precision servedbf16no quantization on this demo specifically (the QLoRA variants load NF4)
training

What was (and wasn't) done here

This page exists purely to give the base checkpoint the same kind of dedicated demo every fine-tuned variant has — nothing below reflects work done by this project on this specific model.

0
training tokens — not pretrained by us
$0.00
training cost — none, this is a direct download
Google
pretraining + instruction-tuning both done upstream
playground

Try it

ready
The completion will stream in here.
honest limitations

What this is (and isn't)

This is google/gemma-2-2b, Google's raw pretrained checkpoint — not gemma-2-2b-it. That distinction matters: an earlier version of this page served gemma-2-2b-it directly with no chat template, expecting it to behave like a base model, but it consistently refused to — even asked to continue an incomplete sentence, it responded in "assistant mode," asking for clarification instead of completing the text. That's a real, verified finding: RLHF/instruction-tuning generalized strongly enough in that checkpoint that removing the chat template alone wasn't sufficient to recover base-model completion behavior. Switching to the actual non-instruction-tuned checkpoint fixed it.

It has no legal/financial specialization and no exposure to the Q&A dataset every fine-tuned Gemma variant here was trained on — and unlike them, it isn't the starting point they were built from (they all start from gemma-2-2b-it). It's shown mainly as a reference point: compare its completions against the SLM-125M base demo (a true from-scratch base model) or against Gemma-2-2B-SFT to see what fine-tuning actually changed.