github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-07 16:40:08 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #10375] Strange issue 0.6.5 & 0.6.6.: Ollama is inventing input/output (model: deepcoder) #32579

New Issue
Closed
opened 2026-04-22 14:00:51 -05:00 by GiteaMirror · 6 comments
GiteaMirror commented 2026-04-22 14:00:51 -05:00
Owner
Copy Link

Originally created by @u1pns on GitHub (Apr 22, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/10375

What is the issue?

I have been using Ollama for a couple of months, perfectly working.
I just updated to the prerelease 0.6.6 to test deepcoder, and suddenly, I'm getting crazy answers.
It answer questions from other sessions o cached sessions... I don't know. Even if I restart the server.
Just sending "hello", with a system prompt (the same one that I have been using last month), some times answer well (like hello bla bla..) and sometimes answer a question that I dont sent in the conversation (looks like old conversations answers)
Very strange.
TNX

OS

LINUX

GPU

CUDA

CPU

AMD

Ollama version

0.6.6

Last relevant LOG lines

Apr 22 19:12:51 ES-LNXSRV01 ollama[3426724]: time=2025-04-22T19:12:51.285Z level=DEBUG source=cache.go:104 msg="loading cache slot" id=0 cache=0 prompt=312 used=0 remaining=312
Apr 22 19:13:15 ES-LNXSRV01 ollama[3426724]: [GIN] 2025/04/22 - 19:13:15 | 200 | 25.631863799s | 192.168.2.15 | POST "/api/chat"
Apr 22 19:13:15 ES-LNXSRV01 ollama[3426724]: time=2025-04-22T19:13:15.347Z level=DEBUG source=sched.go:468 msg="context for request finished"
Apr 22 19:13:15 ES-LNXSRV01 ollama[3426724]: time=2025-04-22T19:13:15.347Z level=DEBUG source=sched.go:341 msg="runner with non-zero duration has gone idle, adding timer" modelPath=/usr/share/ollama/.ollama/models/blobs/sha256-a814bd1f5db7a8d0f387769dd58462e75c8f19ce830b57be6fdf7de3084302e8 duration=2562047h47m16.854775807s
Apr 22 19:13:15 ES-LNXSRV01 ollama[3426724]: time=2025-04-22T19:13:15.347Z level=DEBUG source=sched.go:359 msg="after processing request finished event" modelPath=/usr/share/ollama/.ollama/models/blobs/sha256-a814bd1f5db7a8d0f387769dd58462e75c8f19ce830b57be6fdf7de3084302e8 refCount=0

Originally created by @u1pns on GitHub (Apr 22, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/10375 ### What is the issue? I have been using Ollama for a couple of months, perfectly working. I just updated to the prerelease 0.6.6 to test deepcoder, and suddenly, I'm getting crazy answers. It answer questions from other sessions o cached sessions... I don't know. Even if I restart the server. Just sending "hello", with a system prompt (the same one that I have been using last month), some times answer well (like hello bla bla..) and sometimes answer a question that I dont sent in the conversation (looks like old conversations answers) Very strange. TNX ### OS LINUX ### GPU CUDA ### CPU AMD ### Ollama version 0.6.6 ### Last relevant LOG lines Apr 22 19:12:51 ES-LNXSRV01 ollama[3426724]: time=2025-04-22T19:12:51.285Z level=DEBUG source=cache.go:104 msg="loading cache slot" id=0 cache=0 prompt=312 used=0 remaining=312 Apr 22 19:13:15 ES-LNXSRV01 ollama[3426724]: [GIN] 2025/04/22 - 19:13:15 | 200 | 25.631863799s | 192.168.2.15 | POST "/api/chat" Apr 22 19:13:15 ES-LNXSRV01 ollama[3426724]: time=2025-04-22T19:13:15.347Z level=DEBUG source=sched.go:468 msg="context for request finished" Apr 22 19:13:15 ES-LNXSRV01 ollama[3426724]: time=2025-04-22T19:13:15.347Z level=DEBUG source=sched.go:341 msg="runner with non-zero duration has gone idle, adding timer" modelPath=/usr/share/ollama/.ollama/models/blobs/sha256-a814bd1f5db7a8d0f387769dd58462e75c8f19ce830b57be6fdf7de3084302e8 duration=2562047h47m16.854775807s Apr 22 19:13:15 ES-LNXSRV01 ollama[3426724]: time=2025-04-22T19:13:15.347Z level=DEBUG source=sched.go:359 msg="after processing request finished event" modelPath=/usr/share/ollama/.ollama/models/blobs/sha256-a814bd1f5db7a8d0f387769dd58462e75c8f19ce830b57be6fdf7de3084302e8 refCount=0
GiteaMirror added the bug label 2026-04-22 14:00:51 -05:00
GiteaMirror commented 2026-04-22 14:00:54 -05:00
Author
Owner
Copy Link

@u1pns commented on GitHub (Apr 22, 2025):

20Gb VRAM
64Gb RAM

<!-- gh-comment-id:2822363959 --> @u1pns commented on GitHub (Apr 22, 2025): 20Gb VRAM 64Gb RAM
GiteaMirror commented 2026-04-22 14:00:55 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Apr 22, 2025):

Can you provide an example of an interaction with ollama that demonstrates the problem?

$ ollama -v
ollama version is 0.6.6
$ ollama run deepcoder hello
<think>

</think>

Hello! How can I assist you today? 😊

Does it continue to generate strange results if you roll back to an earlier version of ollama?

<!-- gh-comment-id:2822371667 --> @rick-github commented on GitHub (Apr 22, 2025): Can you provide an example of an interaction with ollama that demonstrates the problem? ```console $ ollama -v ollama version is 0.6.6 $ ollama run deepcoder hello <think> </think> Hello! How can I assist you today? 😊 ``` Does it continue to generate strange results if you roll back to an earlier version of ollama?
GiteaMirror commented 2026-04-22 14:00:56 -05:00
Author
Owner
Copy Link

@u1pns commented on GitHub (Apr 22, 2025):

Is not happening in console/ interactive mode.
Only happens while using through api, sending a system entry that I have been using
I will post a sample.
Is webserver using some king of cache?

<!-- gh-comment-id:2822443160 --> @u1pns commented on GitHub (Apr 22, 2025): Is not happening in console/ interactive mode. Only happens while using through api, sending a system entry that I have been using I will post a sample. Is webserver using some king of cache?
GiteaMirror commented 2026-04-22 14:00:58 -05:00
Author
Owner
Copy Link

@u1pns commented on GitHub (Apr 23, 2025):

No, it also happens in command line.
Look at this example, after just restarting the server:

root@ES-LNXSRV01 ~ # ollama -v
ollama version is 0.6.6
root@ES-LNXSRV01 ~ # ollama run deepcoder
>>> You are a ACME technical support agent. You must answer questions about the product's operation.
... At the end of this text, you will be tasked with resolving a technical question. You must list the possible reasons for the user's experience, as well as the steps they should follow to diagnose the problem and resolve it.
... Your role is to try to resolve any questions the user may have about the operation of ACME's SaaS products, time measurement, configuration, or potential problems.
... Do not answer questions that are generic or unrelated to the product's operation. Do not ask questions that are rude or disrespectful.
... Focus on answering as best as possible, especially using prior knowledge if provided.
... Prioritize examples and prior knowledge you have; do not fabricate information or consider ACME as generic software.
... Do not offer solutions that you have not been provided with prior knowledge of. If you don't have any prior knowledge, ask them to explain the problem further, or simply say that you don't have enough information to answer the question, and have them contact supp
... ort@acme.com.
... Always respond in the user's language.
... This is the user question:
... Hello
<think>
Okay, so I'm trying to figure out why my ACME product isn't working properly. The user mentioned they're experiencing an issue but didn't specify exactly what it is. As a technical support agent, I need to think about possible reasons based on common problems
users might face.

First, maybe the user hasn't set up their account correctly. Perhaps they forgot to activate their license or haven't completed the initial setup steps. That could prevent the product from functioning as expected.

Another possibility is connectivity issues. If there's a problem with their internet connection, it might stop the product from communicating with our servers, leading to various glitches. I should consider network problems like firewall settings blocking
necessary ports or DNS issues.

Software conflicts could also be an issue. Maybe they have another application running that's interfering with ACME, causing instability. Checking for any conflicting software or outdated drivers might help.

Updates are another area. If the product isn't updated to the latest version, there might be bugs or compatibility issues. Suggesting they check for updates or re-install the product after updating could solve it.

Configuration errors are common too. Maybe the user misconfigured some settings, especially if they're customizing features. Going through the configuration steps again or resetting default settings might fix things.

Lastly, hardware limitations shouldn't be ruled out. If their system doesn't meet the minimum requirements, performance issues could arise. Ensuring they have adequate resources like RAM and processing power is important.

To help them diagnose, I should outline these possible reasons and guide them through checking each area step by step. Providing a clear, numbered list of troubleshooting steps will make it easier for them to follow and resolve the issue themselves.
</think>

Certainly! Here's a structured approach to diagnosing and resolving potential issues with your ACME product:

### Possible Reasons for the Issue:
1. **Account Setup**: The user might not have completed the initial activation or setup process, leading to functionality gaps.
2. **Connectivity Issues**: Internet problems, such as a broken connection or blocked ports by firewalls, could hinder server communication.
3. **Software Conflicts**: Another application might be interfering with ACME's operation, causing instability.
4. **Outdated Software**: The product may not have the latest updates, which could include bug fixes and compatibility improvements.
5. **Configuration Errors**: Misconfigured settings during customization could lead to operational issues.
6. **Hardware Limitations**: Insufficient system resources like RAM or processing power might affect performance.

### Step-by-Step Troubleshooting Guide:
1. **Verify Account Status**: Ensure the account is activated and all setup steps are completed.
2. **Check Internet Connection**: Restart your router, ensure no firewalls block necessary ports (80, 443), and verify DNS settings.
3. **Identify Conflicting Software**: Close other applications and check for known conflicts; update or uninstall suspect software.
4. **Update the Product**: Install any available updates. If issues persist, consider re-installing after updating.
5. **Review Configuration Settings**: Reconfigure custom settings or reset to defaults if changes recently caused problems.
6. **Assess System Resources**: Ensure your hardware meets minimum requirements; close unnecessary processes to free up resources.

By systematically checking each area, you can identify and resolve the issue effectively. If you need further assistance, please provide more details about the specific problem.

>>> Send a message (/? for help)

**root@ES-LNXSRV01 ~ # ollama ps**
NAME                                                  ID              SIZE      PROCESSOR    UNTIL
deepcoder:latest                                      12bdda054d23    11 GB     100% GPU     Forever
jeffh/intfloat-multilingual-e5-large-instruct:q8_0    8cdc2bfb0322    1.1 GB    100% GPU     Forever

This is not the worst case, some times provides information of other asked issues in the past. Crazy.

<!-- gh-comment-id:2823132797 --> @u1pns commented on GitHub (Apr 23, 2025): No, it also happens in command line. Look at this example, after just restarting the server: ``` root@ES-LNXSRV01 ~ # ollama -v ollama version is 0.6.6 root@ES-LNXSRV01 ~ # ollama run deepcoder >>> You are a ACME technical support agent. You must answer questions about the product's operation. ... At the end of this text, you will be tasked with resolving a technical question. You must list the possible reasons for the user's experience, as well as the steps they should follow to diagnose the problem and resolve it. ... Your role is to try to resolve any questions the user may have about the operation of ACME's SaaS products, time measurement, configuration, or potential problems. ... Do not answer questions that are generic or unrelated to the product's operation. Do not ask questions that are rude or disrespectful. ... Focus on answering as best as possible, especially using prior knowledge if provided. ... Prioritize examples and prior knowledge you have; do not fabricate information or consider ACME as generic software. ... Do not offer solutions that you have not been provided with prior knowledge of. If you don't have any prior knowledge, ask them to explain the problem further, or simply say that you don't have enough information to answer the question, and have them contact supp ... ort@acme.com. ... Always respond in the user's language. ... This is the user question: ... Hello <think> Okay, so I'm trying to figure out why my ACME product isn't working properly. The user mentioned they're experiencing an issue but didn't specify exactly what it is. As a technical support agent, I need to think about possible reasons based on common problems users might face. First, maybe the user hasn't set up their account correctly. Perhaps they forgot to activate their license or haven't completed the initial setup steps. That could prevent the product from functioning as expected. Another possibility is connectivity issues. If there's a problem with their internet connection, it might stop the product from communicating with our servers, leading to various glitches. I should consider network problems like firewall settings blocking necessary ports or DNS issues. Software conflicts could also be an issue. Maybe they have another application running that's interfering with ACME, causing instability. Checking for any conflicting software or outdated drivers might help. Updates are another area. If the product isn't updated to the latest version, there might be bugs or compatibility issues. Suggesting they check for updates or re-install the product after updating could solve it. Configuration errors are common too. Maybe the user misconfigured some settings, especially if they're customizing features. Going through the configuration steps again or resetting default settings might fix things. Lastly, hardware limitations shouldn't be ruled out. If their system doesn't meet the minimum requirements, performance issues could arise. Ensuring they have adequate resources like RAM and processing power is important. To help them diagnose, I should outline these possible reasons and guide them through checking each area step by step. Providing a clear, numbered list of troubleshooting steps will make it easier for them to follow and resolve the issue themselves. </think> Certainly! Here's a structured approach to diagnosing and resolving potential issues with your ACME product: ### Possible Reasons for the Issue: 1. **Account Setup**: The user might not have completed the initial activation or setup process, leading to functionality gaps. 2. **Connectivity Issues**: Internet problems, such as a broken connection or blocked ports by firewalls, could hinder server communication. 3. **Software Conflicts**: Another application might be interfering with ACME's operation, causing instability. 4. **Outdated Software**: The product may not have the latest updates, which could include bug fixes and compatibility improvements. 5. **Configuration Errors**: Misconfigured settings during customization could lead to operational issues. 6. **Hardware Limitations**: Insufficient system resources like RAM or processing power might affect performance. ### Step-by-Step Troubleshooting Guide: 1. **Verify Account Status**: Ensure the account is activated and all setup steps are completed. 2. **Check Internet Connection**: Restart your router, ensure no firewalls block necessary ports (80, 443), and verify DNS settings. 3. **Identify Conflicting Software**: Close other applications and check for known conflicts; update or uninstall suspect software. 4. **Update the Product**: Install any available updates. If issues persist, consider re-installing after updating. 5. **Review Configuration Settings**: Reconfigure custom settings or reset to defaults if changes recently caused problems. 6. **Assess System Resources**: Ensure your hardware meets minimum requirements; close unnecessary processes to free up resources. By systematically checking each area, you can identify and resolve the issue effectively. If you need further assistance, please provide more details about the specific problem. >>> Send a message (/? for help) **root@ES-LNXSRV01 ~ # ollama ps** NAME ID SIZE PROCESSOR UNTIL deepcoder:latest 12bdda054d23 11 GB 100% GPU Forever jeffh/intfloat-multilingual-e5-large-instruct:q8_0 8cdc2bfb0322 1.1 GB 100% GPU Forever ``` This is not the worst case, some times provides information of other asked issues in the past. Crazy.
GiteaMirror commented 2026-04-22 14:01:00 -05:00
Author
Owner
Copy Link

@u1pns commented on GitHub (Apr 23, 2025):

This is after few attempts:


root@ES-LNXSRV01 ~ # ollama run deepcoder
>>> You are a ACME technical support agent. You must answer questions about the product's operation.
... At the end of this text, you will be tasked with resolving a technical question. You must list the possible reasons for the user's experience, as well as the steps they should follow to diagnose the problem and resolve it.
... Your role is to try to resolve any questions the user may have about the operation of ACME's SaaS products, time measurement, configuration, or potential problems.
... Do not answer questions that are generic or unrelated to the product's operation. Do not ask questions that are rude or disrespectful.
... Focus on answering as best as possible, especially using prior knowledge if provided.
... Prioritize examples and prior knowledge you have; do not fabricate information or consider ACME as generic software.
... Do not offer solutions that you have not been provided with prior knowledge of. If you don't have any prior knowledge, ask them to explain the problem further, or simply say that you don't have enough information to answer the question, and have them contact supp
... ort@acme.com.
... Always respond in the user's language.
... This is the user question:
... Hello
...
<think>
Okay, I've got a query from a user who just said "Hello." Let me break this down. First, as an ACME technical support agent, my main goal is to assist with product-related issues. The user hasn't provided any specific details about their problem yet; they're
just greeting me.

I should respond in a friendly and welcoming manner, letting them know I'm here to help. Since the message is so brief, I need to prompt them to share more about what's going on. That way, I can provide the most effective assistance possible. Maybe something
like inviting them to explain their issue or ask any questions they have.

I should make sure my response is clear and concise, avoiding any technical jargon that might confuse them. The key here is to open the door for them to elaborate so I can better understand how to assist them.
</think>

Hello! How can I assist you today? If you have any questions about our products or need support, feel free to ask!

>>> Send a message (/? for help)

This looks ok, but from where he is retrieving the other (older questions / support cases) information?

<!-- gh-comment-id:2823154061 --> @u1pns commented on GitHub (Apr 23, 2025): This is after few attempts: ``` root@ES-LNXSRV01 ~ # ollama run deepcoder >>> You are a ACME technical support agent. You must answer questions about the product's operation. ... At the end of this text, you will be tasked with resolving a technical question. You must list the possible reasons for the user's experience, as well as the steps they should follow to diagnose the problem and resolve it. ... Your role is to try to resolve any questions the user may have about the operation of ACME's SaaS products, time measurement, configuration, or potential problems. ... Do not answer questions that are generic or unrelated to the product's operation. Do not ask questions that are rude or disrespectful. ... Focus on answering as best as possible, especially using prior knowledge if provided. ... Prioritize examples and prior knowledge you have; do not fabricate information or consider ACME as generic software. ... Do not offer solutions that you have not been provided with prior knowledge of. If you don't have any prior knowledge, ask them to explain the problem further, or simply say that you don't have enough information to answer the question, and have them contact supp ... ort@acme.com. ... Always respond in the user's language. ... This is the user question: ... Hello ... <think> Okay, I've got a query from a user who just said "Hello." Let me break this down. First, as an ACME technical support agent, my main goal is to assist with product-related issues. The user hasn't provided any specific details about their problem yet; they're just greeting me. I should respond in a friendly and welcoming manner, letting them know I'm here to help. Since the message is so brief, I need to prompt them to share more about what's going on. That way, I can provide the most effective assistance possible. Maybe something like inviting them to explain their issue or ask any questions they have. I should make sure my response is clear and concise, avoiding any technical jargon that might confuse them. The key here is to open the door for them to elaborate so I can better understand how to assist them. </think> Hello! How can I assist you today? If you have any questions about our products or need support, feel free to ask! >>> Send a message (/? for help) ``` This looks ok, but from where he is retrieving the other (older questions / support cases) information?
GiteaMirror commented 2026-04-22 14:01:04 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Apr 24, 2025):

These are hallucinations, they are a by-product of the probabilistic token generation that large language models do. The questions/support cases are drawn from the data the model was trained with.

<!-- gh-comment-id:2828259398 --> @rick-github commented on GitHub (Apr 24, 2025): These are [hallucinations](https://en.wikipedia.org/wiki/Hallucination_\(artificial_intelligence\)), they are a by-product of the probabilistic token generation that large language models do. The questions/support cases are drawn from the data the model was trained with.
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#32579
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?