Tags don't work at all, lyrics overwrite prompt

[OGR-3]

Okay, so I am finding firstly that this thing actively refuses to make rock or metal music, and more interestingly, if you try to make a non country song with lyrics about a cowboy, it will overwrite everything and make it into a slow country song with a male vocalist every time. So it is taking prompting from the lyrics and the tags are not working at all, really.

Case in point for reference:

Tags: metal, rock, female vocals, intense, aggressive
Lyrics: (song about a cowboy)
Results: slow, whining country song with male vocals

(repeated four additional times with same results)

Tags: metal, male vocals, intense, aggressive
Lyrics: (song about a cowboy)
Results: slow, whining country song with male vocals

Tags: Rock, intense, aggressive
Lyrics: (Song about a cowboy)
Results: slow, whining country song with male vocals

Tags: Rock, intense, aggressive
Lyrics: (Song about a villain)
Results: slow piano music with a monotonous female reading the lyrics

[OGR-3]

Okay so I noticed the part about not putting a space between tags, took the spaces out and at least now it's kinda following the tags, but please in the name of the spaghetti monster please get some better rock and metal clips to train on for 7B

[OneMonkeyArmy]

Why your "Cowboy" is becoming Country (and how to fix it)

Hi! I saw your struggle with the "Cowboy" lyrics forcing a Country style despite your tags. This is a classic case of Semantic Overwrite where the lyrics' weight crushes the prompt tags.

Based on a deep analysis of the HeartMuLa research paper (Sections 3.2 & 6.2) that I’ve just compiled into a guide, here is why this is happening and how to fix it:

1. The "Anchor" Conflict

In HeartMuLa, Genre has a 95% importance rate, but if you provide multiple genres (like "Metal, Rock"), you create Probability Interference. The model gets "lost" and defaults to the strongest semantic hint it can find: your lyrics. If it sees "Cowboy," it ignores your split tags and goes straight to Country.

2. Semantic Gravity

Words like "Cowboy" have massive "gravity" toward Country music in the training data. To fight this, you need a Single Strong Anchor.

How to fix your prompt (The "Less is More" Strategy):

Instead of listing many conflicting tags, try this structure:

• Tag: Heavy Metal (Choose only ONE genre to lock the anchor).
• Timbre: Distorted (Add this to force the model away from acoustic country sounds).
• Mood: Aggressive (To counter the "slow/whining" default).

Recommended Prompt Format:
Heavy Metal, Distorted, Aggressive, Male Vocals

---

I’ve created a comprehensive Tags Guide based on the model's 8 primary training categories to help with these exact situations. You can find the full logic and the importance percentages here to help stabilize your generations:

👉 https://github.com/OneMonkeyArmy/heartlib/blob/main/TAGS_GUIDE.md

Hope this helps you get that Metal Cowboy track!