[GH-ISSUE #6691] Is everything fine with phi3 model? #4210

Open
opened 2026-04-12 15:08:58 -05:00 by GiteaMirror · 5 comments
Owner

Originally created by @eirnym on GitHub (Sep 7, 2024).
Original GitHub issue: https://github.com/ollama/ollama/issues/6691

What is the issue?

I downloaded model 3 moths ago and it worked fine, but now it doesn't work at all.

My query is generate 20 non-existing random English-sounding nouns, less than 6 sylables. Previously it just generated words without descriptions as expected, now with them.

When I substitute "English" with "Polish", it goes into an infinite loop and when I put "German", it start to spill out UUIDs.

Example of Polish output:

1. Krzeszinski
2. Szmaragdowa
3. Złotyka
4. Pomocnicza
5. Wesołeńca
6. Jędrzejki
7. Kartwinka
8. Chrobotnica
9. Skrępijny
1 end. 20 nouns generated successfully! Now, let's shuffle them:

Shuffled List (Randomized):
4. Pomocnicza
6. Jędrzejki
7. Kartwinka
3. Złotyka
9. Skrępijny
1 end. 20 nouns generated successfully! Now, let's shuffle them:

Shuffled List (Randomized):
(and it repeats forever)

Example of German output:

1. Torgelichtweisenheit
... (8 another words correctly generated)
10. Sonnenfinsternistränenqualm
1de25af6-bb4a-3c17-bf8a-9d6e989e3ecc_GermanSoundingNouns=nonExistingWordsList=[Torgelichtweisenheit,Fuchsbärennachtfrost,Himmelspechvogelzunge,...,Sonnenfinsternistränenqualm1de25af6-bb4a-3c17-bf8a-9d6e989e3ecc_GermanSoundingNouns=nonExistingWordsList=[Torgelichtweisenheit,Fuchsbärennachtfrost,Himmelspechvogelzunge,...,de25af6-bb4a-3c17-bf8a-9d6e989e3ecc]

OS

macOS

GPU

Apple

CPU

Apple

Ollama version

ollama version is 0.3.9
previous ollama version was 0.2.3

logs: ollama.log

Originally created by @eirnym on GitHub (Sep 7, 2024). Original GitHub issue: https://github.com/ollama/ollama/issues/6691 ### What is the issue? I downloaded model 3 moths ago and it worked fine, but now it doesn't work at all. My query is `generate 20 non-existing random English-sounding nouns, less than 6 sylables`. Previously it just generated words without descriptions as expected, now with them. When I substitute "English" with "Polish", it goes into an infinite loop and when I put "German", it start to spill out UUIDs. Example of Polish output: ``` 1. Krzeszinski 2. Szmaragdowa 3. Złotyka 4. Pomocnicza 5. Wesołeńca 6. Jędrzejki 7. Kartwinka 8. Chrobotnica 9. Skrępijny 1 end. 20 nouns generated successfully! Now, let's shuffle them: Shuffled List (Randomized): 4. Pomocnicza 6. Jędrzejki 7. Kartwinka 3. Złotyka 9. Skrępijny 1 end. 20 nouns generated successfully! Now, let's shuffle them: Shuffled List (Randomized): (and it repeats forever) ``` Example of German output: ``` 1. Torgelichtweisenheit ... (8 another words correctly generated) 10. Sonnenfinsternistränenqualm 1de25af6-bb4a-3c17-bf8a-9d6e989e3ecc_GermanSoundingNouns=nonExistingWordsList=[Torgelichtweisenheit,Fuchsbärennachtfrost,Himmelspechvogelzunge,...,Sonnenfinsternistränenqualm1de25af6-bb4a-3c17-bf8a-9d6e989e3ecc_GermanSoundingNouns=nonExistingWordsList=[Torgelichtweisenheit,Fuchsbärennachtfrost,Himmelspechvogelzunge,...,de25af6-bb4a-3c17-bf8a-9d6e989e3ecc] ``` ### OS macOS ### GPU Apple ### CPU Apple ### Ollama version ollama version is 0.3.9 previous ollama version was 0.2.3 logs: [ollama.log](https://github.com/user-attachments/files/16919885/ollama.log)
GiteaMirror added the bug label 2026-04-12 15:08:58 -05:00
Author
Owner

@rick-github commented on GitHub (Sep 7, 2024):

If you haven't updated the model then the variable would be the version of ollama. If you could provide server logs from a version where it worked and a version where it doesn't it may help to debug.

Out of curiosity, I used your prompt to run a test over a range of ollama versions on a Linux/Nvidia system:

$ for i in 0.2.{0..8} 0.3.{0..9} ; do OLLAMA_LLM_LIBRARY=cuda_v11 OLLAMA_DOCKER_TAG=$i docker compose up -d ollama 2>/dev/null ; sleep 3 ;  echo $i $(curl -s localhost:11434/api/generate -d '{"model":"phi3","prompt":"generate 20 non-existing random English-sounding nouns, less than 6 sylables","options":{"seed":1},"stream":false}' | jq -r .response | md5sum) ; done
0.2.0 f69092287bd5817adac7e46e4d778429 -
0.2.1 f69092287bd5817adac7e46e4d778429 -
0.2.2 f69092287bd5817adac7e46e4d778429 -
0.2.3 f69092287bd5817adac7e46e4d778429 -
0.2.4 f69092287bd5817adac7e46e4d778429 -
0.2.5 f69092287bd5817adac7e46e4d778429 -
0.2.6 f69092287bd5817adac7e46e4d778429 -
0.2.7 f69092287bd5817adac7e46e4d778429 -
0.2.8 e478b9e334df4d0a627c8bbc084d471d -
0.3.0 e478b9e334df4d0a627c8bbc084d471d -
0.3.1 1f44b69aa18c2a409520c27dc2e99290 -
0.3.2 1f44b69aa18c2a409520c27dc2e99290 -
0.3.3 1f44b69aa18c2a409520c27dc2e99290 -
0.3.4 1f44b69aa18c2a409520c27dc2e99290 -
0.3.5 1f44b69aa18c2a409520c27dc2e99290 -
0.3.6 1f44b69aa18c2a409520c27dc2e99290 -
0.3.7 1f44b69aa18c2a409520c27dc2e99290 -
0.3.8 1f44b69aa18c2a409520c27dc2e99290 -
0.3.9 1f44b69aa18c2a409520c27dc2e99290 -

There was a change at 0.2.8 and again at 0.3.1, but it's been consistent since then (note I had to override the library because cuda_v12 in 0.3.7+ doesn't work with my old Nvidia driver). The llama.cpp backend was updated in 0.2.8 and 0.3.1 so I expect that's why the output changed.

I did the same test for German and Polish nouns with the same pattern of results: stable 0.2.0-0.2.7, 0.2.8-0.3.0, 0.3.1-0.3.9.

<!-- gh-comment-id:2336425005 --> @rick-github commented on GitHub (Sep 7, 2024): If you haven't updated the model then the variable would be the version of ollama. If you could provide [server logs](https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md#how-to-troubleshoot-issues) from a version where it worked and a version where it doesn't it may help to debug. Out of curiosity, I used your prompt to run a test over a range of ollama versions on a Linux/Nvidia system: ``` $ for i in 0.2.{0..8} 0.3.{0..9} ; do OLLAMA_LLM_LIBRARY=cuda_v11 OLLAMA_DOCKER_TAG=$i docker compose up -d ollama 2>/dev/null ; sleep 3 ; echo $i $(curl -s localhost:11434/api/generate -d '{"model":"phi3","prompt":"generate 20 non-existing random English-sounding nouns, less than 6 sylables","options":{"seed":1},"stream":false}' | jq -r .response | md5sum) ; done 0.2.0 f69092287bd5817adac7e46e4d778429 - 0.2.1 f69092287bd5817adac7e46e4d778429 - 0.2.2 f69092287bd5817adac7e46e4d778429 - 0.2.3 f69092287bd5817adac7e46e4d778429 - 0.2.4 f69092287bd5817adac7e46e4d778429 - 0.2.5 f69092287bd5817adac7e46e4d778429 - 0.2.6 f69092287bd5817adac7e46e4d778429 - 0.2.7 f69092287bd5817adac7e46e4d778429 - 0.2.8 e478b9e334df4d0a627c8bbc084d471d - 0.3.0 e478b9e334df4d0a627c8bbc084d471d - 0.3.1 1f44b69aa18c2a409520c27dc2e99290 - 0.3.2 1f44b69aa18c2a409520c27dc2e99290 - 0.3.3 1f44b69aa18c2a409520c27dc2e99290 - 0.3.4 1f44b69aa18c2a409520c27dc2e99290 - 0.3.5 1f44b69aa18c2a409520c27dc2e99290 - 0.3.6 1f44b69aa18c2a409520c27dc2e99290 - 0.3.7 1f44b69aa18c2a409520c27dc2e99290 - 0.3.8 1f44b69aa18c2a409520c27dc2e99290 - 0.3.9 1f44b69aa18c2a409520c27dc2e99290 - ``` There was a change at 0.2.8 and again at 0.3.1, but it's been consistent since then (note I had to override the library because cuda_v12 in 0.3.7+ doesn't work with my old Nvidia driver). The llama.cpp backend was updated in [0.2.8](https://github.com/ollama/ollama/commit/f8fedbda20b1b2531499ba64758642b0568b6f01) and [0.3.1](https://github.com/ollama/ollama/commit/68ee42f995a04bd163eb1c714f53d4c25ab25474) so I expect that's why the output changed. I did the same test for German and Polish nouns with the same pattern of results: stable 0.2.0-0.2.7, 0.2.8-0.3.0, 0.3.1-0.3.9.
Author
Owner

@eirnym commented on GitHub (Sep 7, 2024):

I've updated both Ollama and model as updated model gave me that some library has not been found (#2386)

I use version from GitHub releases

UPD: I've attached logs to the first message

I don't know if ollama saves logs at all.

<!-- gh-comment-id:2336434911 --> @eirnym commented on GitHub (Sep 7, 2024): I've updated both Ollama and model as updated model gave me that some library has not been found (#2386) I use version from GitHub releases UPD: I've attached logs to the first message I don't know if ollama saves logs at all.
Author
Owner

@eirnym commented on GitHub (Sep 7, 2024):

previous model sometimes can't provide proper responses for Polish languages, but at least tried without any loops or anything else.

<!-- gh-comment-id:2336435253 --> @eirnym commented on GitHub (Sep 7, 2024): previous model sometimes can't provide proper responses for Polish languages, but at least tried without any loops or anything else.
Author
Owner

@pdevine commented on GitHub (Sep 12, 2024):

@eirnym you can see the Troubleshooting FAQ here which includes instructions on how to find the server logs.

<!-- gh-comment-id:2345087105 --> @pdevine commented on GitHub (Sep 12, 2024): @eirnym you can see the [Troubleshooting FAQ here](https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md#how-to-troubleshoot-issues) which includes instructions on how to find the server logs.
Author
Owner

@eirnym commented on GitHub (Sep 12, 2024):

@pdevine thank you for a link to documentation.

In my folder there's no logs. It always had files history, id_ed25519 and id_ed25519.pub as well as models folder. And that's it. I've checked ~/Library as well but there's no trace of Ollama as well.

Journalctl command exists only in Linux

<!-- gh-comment-id:2347233515 --> @eirnym commented on GitHub (Sep 12, 2024): @pdevine thank you for a link to documentation. In my folder there's no logs. It always had files `history`, `id_ed25519` and `id_ed25519.pub` as well as `models` folder. And that's it. I've checked `~/Library` as well but there's no trace of Ollama as well. Journalctl command exists only in Linux
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#4210