Ollama tag openbmb/minicpm5 defaults to top_p 0.8, matching neither documented sampling profile

[laurigates]

What

The openbmb/minicpm5 Ollama tag ships these defaults:

$ curl -s localhost:11434/api/show -d '{"model":"openbmb/minicpm5"}' | jq -r '.parameters'
top_p                          0.8
stop                           "<|im_end|>"
stop                           "<|endoftext|>"
temperature                    0.7

The MiniCPM5-1B model card documents top_p 0.95 for both sampling profiles (Think: temperature 0.9, top_p 0.95; No-Think: temperature 0.7, top_p 0.95).

References: ollama.com/openbmb/minicpm5

Why

temperature 0.7 is consistent with choosing the No-Think profile as the single static default — a reasonable choice, since a Modelfile carries one profile while think is toggleable per request. top_p 0.8 matches neither documented profile, so Ollama users sample outside both recommended configurations, and it is unclear which mode the defaults represent.

How

• If unintentional: update the tag's Modelfile to top_p 0.95, aligning with the documented No-Think profile
• If intentional: document the Ollama-specific default and its rationale on the tag page or deploy guide
• Optional: note in the tag description which profile (Think vs No-Think) the defaults target

Environment: Ollama server 0.30.5 / client 0.30.7, macOS (Apple Silicon), openbmb/minicpm5:latest (q4_K_M, 688 MB).

[zhangtao2-1]

@laurigates
thanks for the precise report — you're right, and this is an oversight on our side.

The intent was to ship the No-Think profile as the static default (hence temperature 0.7), since a Modelfile can only carry one profile while think is toggled per request. The top_p 0.8 doesn't match either documented profile — it's a leftover default and should be 0.95.

We'll update the openbmb/minicpm5 Ollama Modelfile to top_p 0.95 to align with the documented No-Think profile, and add a note on the tag page clarifying that the defaults target No-Think (with think toggleable per request). In the meantime, you can override per request, e.g. "options": {"top_p": 0.95}.

Will update here once the tag is re-pushed. Thanks again for flagging it.