Qwen3.5-4B-EU-Tool

Qwen3.5-4B-EU-Tool

A European, tool-use-capable post-training of Qwen/Qwen3.5-4B. It folds function-calling / agentic data into the EU instruction mix, so the model calls tools in Qwen3.5's native <tool_call> format while keeping European multilingual instruction following. Defaults to fast no-think answers with a /think reasoning toggle. Part of OpenEuroLLM Task 4.6, trained on LUMI (AMD MI250X / ROCm).

Tool-calling evaluation (BFCL-style)

A vLLM-free, self-contained function-calling eval (scripts/eval_tool_calling.py): 300 held-out prompts (150 that require a call + 150 that should not be called), scored on argument-exact match, function-name match, well-formed output, and relevance (correct abstention):

The base is already tool-capable (native support); this post-training's win is argument precision (+8.7 pts args_exact) while improving relevance (knowing when not to call).

A note on RL. A verifiable-reward GRPO stage was also trained, but its tool_call_exact reward had no abstention component, so it reward-hacked — driving relevance to 0.0 (it always calls). We therefore ship the SFT, not the GRPO. A fixed reward (crediting correct abstention, with no-call examples) is the follow-up.

Training data

Openly-documented European post-training data (no proprietary data). SFT mix (~590k), packed, bf16:

• General EU instructions (EuroBlocks EU-multilingual + Tülu-3 English replay), no-think.
• Reasoning traces (chain-of-thought, think-format) — keeps the /think path sharp.
• Defect-repair (native text for weak EU languages), oversampled.
• Tool-use (~10%) — birgermoell/oellm-eu-tooluse-v1: ~55k function-calling examples normalized to Qwen3.5-native format from Glaive-v2 / ToolACE / Hermes (all Apache-2.0).

Usage (with tools)

from transformers import AutoModelForCausalLM, AutoTokenizer  # transformers >= 5.5 (qwen3_5 arch)
import torch

mid = "birgermoell/Qwen3.5-4B-EU-Tool"
tok = AutoTokenizer.from_pretrained(mid)
model = AutoModelForCausalLM.from_pretrained(mid, dtype=torch.bfloat16, device_map="auto")

tools = [{"type": "function", "function": {
    "name": "get_weather",
    "description": "Get the current weather for a city",
    "parameters": {"type": "object",
        "properties": {"city": {"type": "string"}}, "required": ["city"]}}}]
msgs = [{"role": "user", "content": "Vad är vädret i Stockholm?"}]
ids = tok.apply_chat_template(msgs, tools=tools, add_generation_prompt=True,
                             return_tensors="pt").to(model.device)
print(tok.decode(model.generate(ids, max_new_tokens=200)[0, ids.shape[1]:], skip_special_tokens=True))
# -> <tool_call>\n{"name": "get_weather", "arguments": {"city": "Stockholm"}}\n</tool_call>

• No-think (default): fast, direct answers / tool calls.
• Thinking: add /think (or enable_thinking=True) for step-by-step reasoning.

License

Apache-2.0 (inherits the Qwen3.5-4B base license). Built within OpenEuroLLM.