Use this if you want a single from_pretrained() load without PEFT dependencies. Use the adapter-only version if you want to swap/combine adapters.

Performance

Metric	Synthetic heldout	Realistic heldout
Combined accuracy	91.2% (52/57)	98.2% (56/57)
Tool accuracy	98.2% (56/57)	98.2% (56/57)
Args accuracy	91.2% (52/57)	98.2% (56/57)

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL = "Pimmetjeoss/tribe-crm-qwen-2b-merged"
tokenizer = AutoTokenizer.from_pretrained(MODEL)
model = AutoModelForCausalLM.from_pretrained(MODEL, dtype="bfloat16").to("cuda").eval()

messages = [
    {"role": "system", "content": "You are a model that can do function calling with the following functions"},
    {"role": "user", "content": "Zoek Bakkerij De Wit op"},
]
inputs = tokenizer.apply_chat_template(messages, tools=YOUR_TOOLS_SCHEMA,
                                       add_generation_prompt=True,
                                       return_dict=True, return_tensors="pt").to("cuda")
out = model.generate(**inputs, max_new_tokens=128, do_sample=False)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:]))

Adapter-only: Pimmetjeoss/tribe-crm-qwen-2b
GGUF (q4_k_m for llama.cpp/Ollama): Pimmetjeoss/tribe-crm-qwen-2b-gguf
Training pipeline: GitHub
Base: unsloth/Qwen3.5-2B

Pimmetjeoss/tribe-crm-qwen-2b-merged

Tribe CRM Agent — Qwen3.5-2B (merged)

Performance

Usage

Related

Pimmetjeoss/
tribe-crm-qwen-2b-merged