Pimmetjeoss/tribe-crm-qwen-2b

Tribe CRM Agent — Qwen3.5-2B LoRA

LoRA adapter on top of unsloth/Qwen3.5-2B, fine-tuned for Dutch-language tool-calling against the Tribe CRM OData API.

Performance

Evaluated on a held-out set of 57 Dutch CRM prompts:

Metric	Score
Combined accuracy	91.2%
Tool accuracy	98.2%
Args accuracy	91.2%

Compared to FunctionGemma 270M baseline on the same held-out set: combined 91.2% vs 59.6% (+31.6pp).

Fully resolves three known weak tools from prior iterations: zoek_contact_op_email, maak_factuur, voeg_productregel_toe — all 100%.

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

BASE = "unsloth/Qwen3.5-2B"
ADAPTER = "Pimmetjeoss/tribe-crm-qwen-2b"

tokenizer = AutoTokenizer.from_pretrained(ADAPTER)
base = AutoModelForCausalLM.from_pretrained(BASE, dtype="bfloat16")
model = PeftModel.from_pretrained(base, ADAPTER).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]:]))

Output format is Qwen native XML-style tool calls:

<tool_call>
<function=zoek_contact>
<parameter=naam>
Bakkerij De Wit
</parameter>
</function>
</tool_call>

A parser for this format is available at Pimmetjeoss/tribe-crm-training → agent/parser.py.

Training

• Base: unsloth/Qwen3.5-2B
• Method: QLoRA 4-bit via Unsloth Studio
• Dataset: 978 Dutch Tribe CRM tool-calling examples (tribe_crm_dataset_v2 + hard_negatives + multiturn_observations)
• LoRA: rank 16, alpha 16, all 7 linear modules targeted
• Schedule: 1 epoch, cosine LR, warmup 5%
• Hardware: RTX 3060 12GB

Related

• Base model: unsloth/Qwen3.5-2B
• Training pipeline: Pimmetjeoss/tribe-crm-training
• Previous FunctionGemma baseline: Pimmetjeoss/tribe-crm-functiongemma-270m