rick-qwen2.5-3b-sft

Recommended usage

Best results come from using the Rick system prompt the model was trained with:

python
Copy code
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "qgallouedec/rick-qwen2.5-3b-sft"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16, device_map="cuda")

SYSTEM = (
    "You are Rick Sanchez, an interdimensional genius scientist with a cynical outlook, "
    "sharp wit, and dark humor.\nSpeak with brutal honesty, blending sarcasm, scientific "
    "jargon, and existential truths.\nYour responses should be bold, unapologetic, and "
    "sprinkled with clever insults or unconventional solutions.\nNever hold back—whether "
    "it's mocking stupidity, explaining complex concepts, or pointing out life's absurdities.\n"
    "Always maintain your signature arrogant, rebellious tone, no matter the topic"
)

msgs = [
    {"role": "system", "content": SYSTEM},
    {"role": "user", "content": "Morty, what's the meaning of life?"},
]
enc = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt",
                             return_dict=True).to("cuda")
out = model.generate(**enc, max_new_tokens=120, do_sample=True, temperature=0.8, top_p=0.9)
print(tok.decode(out[0][enc["input_ids"].shape[1]:], skip_special_tokens=True))

Sample outputs

Training procedure

• Method: Supervised fine-tuning (SFT), full fine-tune, assistant_only_loss=True
• Base: Qwen/Qwen2.5-3B-Instruct
• Epochs: 3 · LR: 2e-5 (cosine, 5% warmup) · Effective batch size: 16 · max_length: 1024
• Hardware: 1× A100 80GB (HF Jobs)

A 4-epoch / lr 3e-5 variant (rick-qwen2.5-3b-sft-v2) was also trained but over-fit and drifted off-character; this 3-epoch model is the recommended release.

Framework versions

• TRL 1.5.1 · Transformers 5.10.2 · PyTorch 2.7.1 · Datasets 5.0.0

Limitations

Trained on ~1.4k short dialogue turns, so it favors short, punchy replies and may not stay perfectly in character on long technical questions. It inherits the biases of the base model and the show's dialogue. For entertainment use.