github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-07 00:22:43 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #12559] Non-deterministic structured output with same seed #70390

New Issue
Open
opened 2026-05-04 21:23:29 -05:00 by GiteaMirror · 3 comments
GiteaMirror commented 2026-05-04 21:23:29 -05:00
Owner
Copy Link

Originally created by @ndido98 on GitHub (Oct 10, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/12559

What is the issue?

I'm running into an issue in which the output is not deterministic between runs when I ask for structured output instead of plaintext.

See following Python script:

import ollama
from pydantic import BaseModel

class RandomNumber(BaseModel):
    value: int

class RandomNumbers(BaseModel):
    numbers: list[RandomNumber]

response = ollama.chat(
    model="phi4:14b",
    messages=[
        {"role": "user", "content": "Generate 100 random numbers between 1 and 100"}
    ],
    format=RandomNumbers.model_json_schema(),
    options=ollama.Options(
        seed=42,
        temperature=1.0,
        num_gpu=9999,
        top_p=0.95,
        top_k=40,
    ),
)
model = RandomNumbers.model_validate_json(response.message.content)
print([n.value for n in model.numbers])

If I run this script twice, I expect to have the same list of numbers, but instead I get two very different ones. However, if I comment out the format keyword argument and instead just print out the raw model's response, that one is deterministic and is consistently repeated across runs, as expected. In other words, it looks like that the structured output somehow breaks the determinism I expected.

OS

Linux

GPU

Nvidia

CPU

AMD

Ollama version

0.12.3

Originally created by @ndido98 on GitHub (Oct 10, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/12559 ### What is the issue? I'm running into an issue in which the output is not deterministic between runs when I ask for structured output instead of plaintext. See following Python script: ```python import ollama from pydantic import BaseModel class RandomNumber(BaseModel): value: int class RandomNumbers(BaseModel): numbers: list[RandomNumber] response = ollama.chat( model="phi4:14b", messages=[ {"role": "user", "content": "Generate 100 random numbers between 1 and 100"} ], format=RandomNumbers.model_json_schema(), options=ollama.Options( seed=42, temperature=1.0, num_gpu=9999, top_p=0.95, top_k=40, ), ) model = RandomNumbers.model_validate_json(response.message.content) print([n.value for n in model.numbers]) ``` If I run this script twice, I expect to have the same list of numbers, but instead I get two very different ones. However, if I comment out the format keyword argument and instead just print out the raw model's response, that one is deterministic and is consistently repeated across runs, as expected. In other words, it looks like that the structured output somehow breaks the determinism I expected. ### OS Linux ### GPU Nvidia ### CPU AMD ### Ollama version 0.12.3
GiteaMirror added the bug label 2026-05-04 21:23:29 -05:00
GiteaMirror commented 2026-05-04 21:23:30 -05:00
Author
Owner
Copy Link

@ghmer commented on GitHub (Oct 10, 2025):

Isn't the temperature also controlling how "random" a generated answer is?
I was thinking of a seed as "this is the starting point for the generator", temperature == 0 is deterministic, and every temperature over 0 adds some "randomness" to the output.

<!-- gh-comment-id:3390467085 --> @ghmer commented on GitHub (Oct 10, 2025): Isn't the temperature also controlling how "random" a generated answer is? I was thinking of a seed as "this is the starting point for the generator", temperature == 0 is deterministic, and every temperature over 0 adds some "randomness" to the output.
GiteaMirror commented 2026-05-04 21:23:31 -05:00
Author
Owner
Copy Link

@ndido98 commented on GitHub (Oct 10, 2025):

In my mental model the seed should also affect the sampler, so by fixing the seed I should not always get the most probable token (as temperature > 0); however, across different calls I expect the same tokens in the same order if I'm using the same seed, as the sampler should choose the same random tokens.

EDIT: I also tried explicitly setting temperature to 0, but still two different runs yield different responses. However, after the second run, all responses are the same for a while; I suppose I am hitting some sort of cache.

<!-- gh-comment-id:3390647808 --> @ndido98 commented on GitHub (Oct 10, 2025): In my mental model the seed should also affect the sampler, so by fixing the seed I should not always get the most probable token (as temperature > 0); however, across different calls I expect the same tokens in the same order if I'm using the same seed, as the sampler should choose the same random tokens. EDIT: I also tried explicitly setting temperature to 0, but still two different runs yield different responses. However, after the second run, all responses are the same for a while; I suppose I am hitting some sort of cache.
GiteaMirror commented 2026-05-04 21:23:31 -05:00
Author
Owner
Copy Link

@ndido98 commented on GitHub (Oct 10, 2025):

Probably related to #5321 and #5760?

<!-- gh-comment-id:3391182716 --> @ndido98 commented on GitHub (Oct 10, 2025): Probably related to #5321 and #5760?
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
amd
api
app
bug
build
cli
cloud
compatibility
context-length
create
docker
documentation
embeddings
feature request
feedback wanted
good first issue
gpt-oss
gpu
harmony
help wanted
image
install
intel
js
launch
linux
macos
memory
mlx
model
needs more info
networking
nvidia
ollama.com
performance
pull-request
Mirrored from GitHub Pull Request
 python
question
registry
rendering
thinking
tools
top
vulkan
windows
wsl
No Label bug
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
GiteaMirror ninjasurge
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#70390
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?