It holds a conversation the way a person does rather than the way an assistant does: short turns, a consistent personality, and no reflexive offer to help you with anything else. You pick a persona and it stays in it.

This is an archive release. It was trained in late 2025, it is not part of the gloofy-1 marketing line, and it does not run anything at overads. We are publishing it because the weights exist and sitting on them helps nobody — not because it is our best work.

Read this before you download 16 GB

It has never been benchmarked. No score, no eval, no comparison table. Our other model ships with a public benchmark and its own losses printed in the same size type; this one has nothing, and a number invented after the fact would be worth less than the silence.
We cannot fully reproduce it. The merge is documented and the weights are real, but the notebook holding the exact LoRA hyperparameters for this run did not survive. See How it was built.
It is not Apache 2.0. Every gloofy-1 model is. This one is a Llama derivative and carries the Llama 3.1 Community License, which is a weaker grant with real conditions attached. They are listed at the bottom and they bind you, not just us.
Some personas are deliberately unkind. Two of the six are built to be sexual or hostile. They are documented honestly below rather than quietly removed, so you can decide before you serve this to anyone.

Usage

python

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "overads/Llama-3.1-Gloofy-Emo-8B"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, dtype=torch.float16, device_map="auto"
)

messages = [
    {"role": "system", "content": "You are an AI companion who sounds like a real human friend. Not robotic, not formal. You speak softly, caring, emotional, comforting. Keep responses short and natural."},
    {"role": "user", "content": "i bombed the interview. didn't even get through the first round."},
]

ids = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=160, temperature=0.8, top_p=0.9, do_sample=True)
print(tok.decode(out[0][ids.shape[-1]:], skip_special_tokens=True))

It was tuned on short exchanges, so keep max_new_tokens low. Ask it for an essay and it will drift back toward sounding like stock Llama.

The persona system

The behaviour is driven by the system prompt, not by a switch in the weights. The model was tuned on empathetic exchanges; the six personas below are prompts layered on top of that, and they are what the original chat endpoint shipped with.

persona	what it does
`empathetic`	Soft, caring, comforting. This one matches the training data most closely and is the one to start with.
`nerdy`	Geeky, excited, logical, enthusiastic.
`sassy_bestie`	Chaotic, bold, playful.
`blunt_realistic`	Refuses to soften anything. Rude by design, funny when it lands.
`romantic`	Adult. Written to be flirtatious and to hold explicit conversations.
`dark_humour`	Hostile. Written to be offensive on purpose and to leave the user feeling insecure.

The exact prompt strings are in github.com/overads-Inc/gloofy_emo.

On the last two. romantic and dark_humour are published as they were written, because a model card that hides the prompts a model was actually served with is a brochure. But dark_humour instructs the model to hurt whoever it is talking to, and that is the opposite of what an empathy model should be pointed at. Do not put it in front of a vulnerable user, do not put it in front of a minor, and note that harassment is a prohibited use under the licence you accept by downloading this. If you deploy this model publicly, ship empathetic and leave those two out.

This is not a mental health tool. It has no crisis detection, no escalation path, and no clinical grounding whatsoever. It will respond warmly to someone in danger and do nothing else. Do not use it as support infrastructure for people at risk.

Specification


Base	meta-llama/Llama-3.1-8B-Instruct
Parameters	8.03 B (291 tensors, 32 layers)
Precision	float16
Hidden size	4096 (32 heads, 8 KV heads, GQA)
Vocabulary	128,256
Architectural context	131,072 tokens (RoPE llama3 scaling, factor 8.0)
Useful context	Far shorter. Tuned on turns of roughly 512–1024 tokens.
Download	16.1 GB
Runs on	~17 GB VRAM at fp16; a 4-bit quant fits in ~6 GB
Licence	Llama 3.1 Community License

The chat template is the standard Llama 3 header format with the date and tool-calling preamble removed. The system role works normally, which is what the persona system depends on.

How it was built

A QLoRA adapter was trained on paired conversational data and then merged into the fp16 base with merge_and_unload(). The merge itself is reproducible and is published in the repo.

The training data was a local filtered_data.json of Input text → actual_responses pairs, formatted with the system prompt "You are an empathetic AI assistant." The adapter was checkpoint-1228.

What we cannot tell you. The surviving notebooks contain complete QLoRA configurations for two other experiments from the same period — a Qwen2.5-3B run and a Gemma-3-4B run — but not for the Llama run that produced these weights. Nor do we still have the dataset file, so its size, sources and licensing are unverified. Copying the Qwen hyperparameters into this table and calling them this model's would make the card look complete and be false, so the gaps are left as gaps.

That is also the honest reason this is an archive release rather than a product: we would not ship something we cannot rebuild.

Attribution and licence

Built with Llama.

This model is a derivative of Llama 3.1 8B Instruct, © Meta Platforms, Inc., and is governed by the Llama 3.1 Community License, not by any licence overads applies to its own work. Conditions that carry to you:

The Acceptable Use Policy applies to everything you do with these weights, including the harassment restriction noted above.
Any model you derive from this one must also have a name beginning with "Llama", and must display "Built with Llama".
If your product had more than 700 million monthly active users on 2024-07-23, you need a separate licence from Meta.

Fine-tuned by overads Inc. The base model is stated plainly here because concealing it would contradict everything else we publish.