Update Lotus Bot custom response llm model to an uncensored version to prevent discouraging restrictions #5

New Issue

2026-04-20T00:16:08-04:00

jared commented

2026-04-20 00:16:08 -04:00

The custom user responses for User ID: @wynter:mozilla.org and @jared:matrix.lotusguild.org need the model updated to use a llama model that is uncensored but small 1-3b max. llama2-uncensored:latest is a 7b model which is too slow. This change is due to the consistant response of "I cannot provide a response that may be seen as negative or discouraging. Would you like to ask another question?"

The custom user responses for User ID: @wynter:mozilla.org and @jared:matrix.lotusguild.org need the model updated to use a llama model that is uncensored but small 1-3b max. **llama2-uncensored:latest** is a 7b model which is too slow. This change is due to the consistant response of "_I cannot provide a response that may be seen as negative or discouraging. Would you like to ask another question?_"

jared commented

2026-04-20 00:20:57 -04:00

dolphin-phi: 2.7B uncensored Dolphin model by Eric Hartford, based on the Phi language model by Microsoft Research.

sadiq-bd/llama3.2-1b-uncensored: An abliterated version of Llama 3.2 1B Instruct. It is designed to follow instructions without the standard refusals found in the original Meta model [9].
huihui_ai/gemma3-abliterated:1b: A proof-of-concept version of Google's Gemma 3 1B that uses abliteration to bypass safety guardrails [6].
carsenk/llama3.2_1b_2025_uncensored_v2: A community-contributed version often cited in forums for its lack of restrictions [23].

[dolphin-phi](https://ollama.com/library/dolphin-phi): 2.7B uncensored Dolphin model by Eric Hartford, based on the Phi language model by Microsoft Research. [sadiq-bd/llama3.2-1b-uncensored](https://ollama.com/sadiq-bd/llama3.2-1b-uncensored): An abliterated version of Llama 3.2 1B Instruct. It is designed to follow instructions without the standard refusals found in the original Meta model [9]. [huihui_ai/gemma3-abliterated:1b](https://ollama.com/huihui_ai/gemma3-abliterated:1b): A proof-of-concept version of Google's Gemma 3 1B that uses abliteration to bypass safety guardrails [6]. [carsenk/llama3.2_1b_2025_uncensored_v2](https://ollama.com/search?q=uncenso): A community-contributed version often cited in forums for its lack of restrictions [23].

jared commented

2026-04-20 00:28:53 -04:00

Updated to dolphin-phi, testing now

jared commented

2026-04-20 00:32:47 -04:00

also downloaded sadiq-bd/llama3.2-1b-uncensored if we want to test that instead. Also need to fix it so when a user refrences "I" it should be in context of the user not the model.

jared commented

2026-04-20 10:57:44 -04:00

Completed

Switched 8-ball LLM to sadiq-bd/llama3.2-1b-uncensored:latest and fixed both reported issues:

All-caps responses — Added _normalize_caps() post-processing: if >60% of alphabetic chars in the response are uppercase, the response is converted to sentence case automatically. This is a safety net that does not touch normal mixed-case responses.

First-person pronoun confusion — Added _replace_first_person() pre-processing: before the question reaches the model, all first-person pronouns are substituted with the actual sender’s name (Jared or Wynter). Handles: I, me, my, myself, mine, I’m, I’ve, I’ll, I’d.

Examples of pronoun rewriting:

Should I go to the park? → Should Jared go to the park?
Will this help me? → Will this help Jared?
Is this my fault? → Is this Wynter’s fault?

Bot restarted and tested live against Ollama. All changes in /opt/matrixbot/commands.py on LXC 151.

## Completed Switched 8-ball LLM to `sadiq-bd/llama3.2-1b-uncensored:latest` and fixed both reported issues: **All-caps responses** — Added `_normalize_caps()` post-processing: if >60% of alphabetic chars in the response are uppercase, the response is converted to sentence case automatically. This is a safety net that does not touch normal mixed-case responses. **First-person pronoun confusion** — Added `_replace_first_person()` pre-processing: before the question reaches the model, all first-person pronouns are substituted with the actual sender’s name (`Jared` or `Wynter`). Handles: `I`, `me`, `my`, `myself`, `mine`, `I’m`, `I’ve`, `I’ll`, `I’d`. Examples of pronoun rewriting: - `Should I go to the park?` → `Should Jared go to the park?` - `Will this help me?` → `Will this help Jared?` - `Is this my fault?` → `Is this Wynter’s fault?` Bot restarted and tested live against Ollama. All changes in `/opt/matrixbot/commands.py` on LXC 151.

jared closed this issue

2026-04-20 10:57:52 -04:00

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: LotusGuild/matrix#5