Qwen3.5-4B-EU

A European post-training of Qwen/Qwen3.5-4B, specialised for the OpenEuroLLM target languages. The strongest model in our compact EU lineup (2B + 4B). Defaults to fast no-think answers and keeps a /think reasoning toggle.

Built with a reproducible SFT → SimPO → GRPO pipeline on European multilingual data.

Evaluation

OpenEuroLLM EU eval holdouts (38 European languages × 10 task buckets, deterministic scoring), confirmed on the large test_public split (6080 rows) — not just the noisier dev set:

This repo ships the v3 weights — the v2 recipe plus a defect-repair SFT fold-in that fixes the previously-weak European languages. On the holdouts the once-broken languages are now at parity with the rest (defect-language mean 60.3 vs 58.2 for the others; e.g. is 56, mt 70, da 73, bg 68), with no collapsed language across all 38 — and clean task buckets (summarization 93, no_answer 84, reasoning-math 96). The v2/v3 recipe helps at 4B scale but overfit at 2B, so Qwen3.5-2B-EU keeps its lighter recipe.

Training data

All stages use openly-documented European post-training data (OpenEuroLLM Task 4.6); no proprietary data. Three stages: SFT → SimPO → GRPO (the "v2" recipe — more data + longer training + a verifiable-RL stage; this gave a clear, robust gain at 4B scale).

1. Supervised fine-tuning (SFT) — full mix (~1.27M), packed, bf16:

• General EU instructions (~85%, full ~1.08M) — Dolci tulu3-euroblocks-85-15: EuroBlocks EU-multilingual instruction data (85%) + Tülu-3 (allenai/tulu-3) English replay (15%), adding EU-language instruction-following while preserving English. No-think formatting (default fast answers).
• Reasoning traces (~15%, ~190k) — chain-of-thought SFT (Dolci-Think / OpenThoughts / OpenMathInstruct-family), think-format, keeping the /think path sharp. (Currently English+Finnish reasoning — the known multilingual-reasoning gap; EU-language reasoning distillation is in progress.)

2. Preference optimisation (SimPO, reference-free) — preference accuracy ≈ 0.88:

• birgermoell/oellm-eu-exam-mcq-v1 (preference/DPO track) — European exam multiple-choice preference pairs (correct-over-incorrect), ~35 languages, 28 sources (national, medical/licensing, academic exams), mixed licenses.

3. Verifiable RL (GRPO / RLVR) — reward climbed 0.78 → 0.85:

• birgermoell/oellm-eu-exam-mcq-v1 (GRPO/RLVR track) — verifiable letter-match reward (mcq_letter_exact, β=0, no reward model), optimizing EU-language answer correctness directly (no translationese in the loop).

Evaluation — held out from training: birgermoell/oellm-eu-eval-holdouts-v1 — 38 EU languages × 10 task buckets, deterministic per-task scoring.

Swedish capability (open-ended rubric)

A vLLM-free Swedish eval (Qwen3.5 isn't yet supported by EuroEval's fast backends): 24 prompts across 9 categories (sentiment, Swedish knowledge, reasoning, summarization, linguistic correctness, instruction-following, common sense, creative), scored 1–5 on språkkvalitet / korrekthet / instruktion / hjälpsamhet, plus automatic langdetect Swedish-purity.

The 4B is a big step up from the 2B on factual accuracy (korrekthet 2.7 → 4.4): it gets the three largest Swedish lakes right (Mälaren, not Göta älv), explains the 4 % riksdag threshold, knows Selma Lagerlöf won the Nobel, and handles de/dem correctly — all while writing fully in-Swedish (purity 1.00). One wart: a constrained "list five cities" task degenerated into a repetition loop. Tooling: scripts/eval_swedish_rubric.py + data/swedish_rubric_prompts.jsonl.

Usage

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

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

msgs = [{"role": "user", "content": "Vad är meningen med livet? Svara kort."}]
ids = tok.apply_chat_template(msgs, 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))

• No-think (default): fast, direct answers — best for interactive use.
• Thinking: add /think to the user turn (or enable_thinking=True) for step-by-step reasoning.

On-device status

• bf16 (this repo): runs via transformers / vLLM. Recommended today.
• GGUF (llama.cpp / Ollama): ⚠️ not supported yet — llama.cpp drops Qwen3.5's hybrid linear-attention (Gated Delta Net) layers during conversion (load fails). Blocked upstream.
• MLX (Apple): ⚠️ not supported yet — mlx-lm raises Model type qwen3_5 not supported.

Qwen3.5's hybrid architecture is very new (Mar 2026); the on-device quantization runtimes haven't added support yet. Phone deployment works once llama.cpp / MLX add Gated-Delta-Net support.

License

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