MANGO1.5-Qwen3.5-9B

MANGO (Multimodal Adaptation for thaNGuage Optimization) is CMKL University's Thai-language fine-tuning effort. MANGO1.5-Qwen3.5-9B is the newest release in the line, a dense 9B text-only bilingual (Thai/English) fine-tune of Qwen/Qwen3.5-9B, produced via full-model LoRA supervised fine-tuning on a 5,000,000-sample English/Thai instruction corpus.

It supersedes the previous release, CMKL/MANGO-Qwen3-Omni-30B-A3B-Instruct (a 30B MoE omni-modal fine-tune), with a different design point: a smaller, dense, text-only base trained on a substantially larger and more broadly-sourced bilingual SFT corpus (5M vs. ~108K samples). This checkpoint is stage 1 (SFT) of a two-stage post-training plan; a preference-optimization stage (DPO) targeting Thai response quality specifically is planned as a follow-up and is not part of this release.

Model Details

• Base model: Qwen/Qwen3.5-9B — a dense 9B-parameter model from the Qwen3.5 family.
• Fine-tuning method: LoRA (rank 64, alpha 128, dropout 0.05, target_modules=all-linear), applied across the full language model (no frozen sub-modules). Weights in this repository are the LoRA adapter merged into the base weights and exported to standard HF safetensors — this is a complete, standalone checkpoint, not an adapter to be applied separately.
• Training data: 5,000,000 samples (65% English / 35% Thai), text-only instruction-tuning data spanning chat, math, code, STEM, knowledge, creative writing, summarization, translation, safety, legal, medical, and exam domains. See Training Data below for the full domain/language composition; source-level dataset provenance is not published with this release.
• Modality: text input, text output. The base model's architecture (Qwen3_5ForConditionalGeneration) supports multimodal input per its own model card, but this fine-tune was performed on a text-only corpus — no vision data was used and no vision-related behavior was targeted, evaluated, or should be assumed improved here.
• Languages: Thai, English (mixed/code-switched input handled by the same filtering discipline used to build the training data — see below).
• License: inherits the Apache 2.0 license of the base model. See Qwen/Qwen3.5-9B for exact terms.

Intended Use

This model is intended for:

• Thai-language and English-language chat, instruction-following, reasoning, summarization, translation, and Q&A.
• Bilingual use cases where a single model needs to respond fluently in whichever of the two languages the user writes in.

It is trained with a default system prompt that establishes this bilingual persona (see How to Use). It is a research/engineering artifact from CMKL University, not a production-hardened or safety-certified assistant — it inherits the base model's general limitations (hallucination, non-factual output, sensitivity to prompt phrasing), and has only the safety-domain SFT data folded in at this stage (no dedicated safety RLHF pass yet).

How to Use

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "CMKL/MANGO1.5-Qwen3.5-9B"
model = AutoModelForCausalLM.from_pretrained(model_id, dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_id)

system_prompt = (
    "You are Mango, a helpful AI assistant developed by CMKL University that specialized in "
    "Thai and English. You must respond in the language used by the user. If the user "
    "communicates in Thai, provide a culturally nuanced, polite, and grammatically correct "
    "response. If the user communicates in English, maintain high-quality, professional "
    "performance.\n\n"
    "คุณคือ Mango ผู้ช่วย AI ที่เชี่ยวชาญทั้งภาษาไทยและภาษาอังกฤษ ที่พัฒนาโดย มหาวิทยาลัยซีเอ็มเคแอล "
    "(CMKL University) โปรดตอบด้วยภาษาเดียวกับที่ผู้ใช้สื่อสารมา หากผู้ใช้ถามเป็นภาษาไทย "
    "ให้ตอบด้วยภาษาไทยที่สอดคล้องกับบริบททางวัฒนธรรม สุภาพ และถูกต้องตามหลักไวยากรณ์"
)

messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": "อธิบายความแตกต่างระหว่างฝนกับน้ำค้างให้ฟังหน่อย"},
]

inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
output = model.generate(inputs, max_new_tokens=512)
print(tokenizer.decode(output[0][inputs.shape[-1]:], skip_special_tokens=True))

This system prompt is only a fallback default used during training — rows that already carried their own system message in the training data kept it, so you're free to supply a different system prompt at inference time.

Serving with vLLM

vLLM is the recommended way to serve this checkpoint — it's the runtime this model was validated on, including the speculative-decoding path below. Requires a vLLM build with Qwen3.5 hybrid-attention (Gated DeltaNet) support (validated on 0.24.0–0.27.1).

vllm serve CMKL/MANGO1.5-Qwen3.5-9B \
  --port 8000 --tensor-parallel-size 1 --max-model-len 262144 \
  --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder

This checkpoint includes the base model's Multi-Token-Prediction (MTP/NextN) draft layer (see Model Details), so self-speculative decoding can be enabled directly, with no separate draft model:

vllm serve CMKL/MANGO1.5-Qwen3.5-9B \
  --port 8000 --tensor-parallel-size 1 --max-model-len 262144 \
  --reasoning-parser qwen3 \
  --speculative-config '{"model": "CMKL/MANGO1.5-Qwen3.5-9B", "num_speculative_tokens": 3, "method": "mtp"}'

We've validated "method": "mtp" end-to-end (64–76% draft-token acceptance in our own smoke tests); the upstream Qwen3.5-9B base model card documents "method": "qwen3_next_mtp" instead — the accepted method string has moved between vLLM versions, so try mtp first and fall back to qwen3_next_mtp if your vLLM build rejects it.

Serving with SGLang

SGLang also supports Qwen3.5's architecture (per the base model card); we have not personally validated SGLang against this fine-tune (only vLLM, above), but the same recipe should apply directly since the architecture and weights are unchanged from base:

python -m sglang.launch_server --model-path CMKL/MANGO1.5-Qwen3.5-9B \
  --port 8000 --tp-size 1 --mem-fraction-static 0.8 --context-length 262144 \
  --reasoning-parser qwen3 --tool-call-parser qwen3_coder

With MTP speculative decoding:

python -m sglang.launch_server --model-path CMKL/MANGO1.5-Qwen3.5-9B \
  --port 8000 --tp-size 1 --mem-fraction-static 0.8 --context-length 262144 \
  --reasoning-parser qwen3 \
  --speculative-algo NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4

Training Data

5,000,000 samples total, English/Thai bilingual, text-only instruction-tuning data. Composition:

Language ratio

Language	Samples	% of total
English	3,250,000	65.0%
Thai	1,750,000	35.0%

Domain ratio (% of full 5,000,000)

Domain	% of total
chat	27.50%
math	12.04%
code	9.76%
summarization	9.84%
stem	9.68%
creative	6.06%
knowledge	5.53%
translation	5.13%
instruction_following	4.15%
safety	4.11%
medical	2.66%
legal	1.25%
finance	0.69%
tool_calling	0.59%
wiki_qa	0.87%
customer_service	0.10%
exam	0.04%

Some domains are English-only (STEM, knowledge, finance, instruction-following, tool-calling) or Thai-only (translation, legal, medical, wiki QA, customer service, exam) by construction — the source pools for those domains only existed in one of the two languages at adequate quality/volume.

Data preparation

• Language-purity filtering: every row was scored for script composition (Thai-Unicode, Latin, other) and language-ID confidence before being allocated to its language bucket. Rows dominated by a third script (CJK, Cyrillic, Arabic, etc.) were rejected outright. Genuine bilingual/translation pairs are deliberately preserved in the Thai bucket rather than treated as contamination.
• Deduplication: global exact-duplicate removal via content hashing before allocation.
• Domain-balanced allocation: a proportional water-filling allocation targets a fixed domain mix per language without exceeding any domain's real post-filter data pool, and caps how much of a single domain's final allocation any one source can supply, to avoid one large or templated source dominating a domain.
• Benchmark-contamination audit: every candidate source was checked for test/dev/validation splits before inclusion, and known public benchmark sets were excluded from training entirely.

Source-level dataset provenance (which repositories contributed which rows) is intentionally not disclosed with this release.

Training Details

Framework	ms-swift v4.4.2, Megatron-SWIFT backend (Megatron-core 0.18.2), mcore-bridge for direct HF-format weight loading/export
Hardware	1 node × 8× NVIDIA A100 40GB
Parallelism	Tensor parallel = 4, Data parallel = 2, Pipeline parallel = 1, sequence parallel enabled
LoRA config	rank 64, alpha 128, dropout 0.05, target_modules=all-linear, full language model (no frozen sub-modules)
Optimizer	Megatron Adam, LR 1e-4, 5% linear warmup, min LR 1e-5, weight decay 0.1, betas (0.9, 0.95)
Batch size	global 128, micro-batch 4/device, sequence packing enabled
Sequence length	max 8,192 tokens (packed; longer samples dropped, not truncated)
Epochs	3
Precision / kernels	bf16 compute, Flash Attention, full uniform activation recomputation
Wall-clock	~9 days on the above hardware
Final eval loss	0.685 (held-out split, end of epoch 3)

Evaluation

Evaluated on the ThaiLLM Leaderboard methodology — 9 datasets across three categories — against its own base model (Qwen/Qwen3.5-9B) and against Typhoon2.1-Gemma3-12B, a well-established 12B Thai-LLM effort from SCB10X. All rows were scored by the same harness with identical prompts per dataset; the Mango/base rows share an identical serving setup (vLLM, bf16, greedy decoding, 16,384-token context budget), while Typhoon2.1-Gemma3-12B was served with its own repo's recommended chat template and default decoding settings, so minor serving differences there are possible. The fine-tune was evaluated with reasoning toggled on and off, since Qwen3.5's hybrid-thinking template supports both.

Mango1.5 (think mode) has the highest normalized average (0.444) of the four — ahead of Typhoon2.1-Gemma3-12B (0.349) despite being a smaller 9B model, and far ahead of its own un-tuned base (0.305). Typhoon2.1-Gemma3-12B does still lead on two individual datasets: Wisesight sentiment (47.0% vs. Mango's 45.5%) and Flores200 BLEU (38.3 vs. 37.0).

Category	Dataset	Mango (think)	Mango (no-think)	Qwen3.5-9B base¹	Typhoon2.1-Gemma3-12B
Exam	ThaiExam (accuracy)	60.0%	35.2%	30.0%	44.5%
Exam	M3Exam-Thai (accuracy)	65.4%	55.4%	21.6%	18.4%
NLU	Belebele (accuracy)	80.0%	72.1%	58.1%	62.8%
NLU	XNLI (accuracy)	54.7%	57.7%	43.7%	50.4%
NLU	XCOPA (accuracy)	86.6%	73.0%	53.8%	67.4%
NLU	Wisesight sentiment (accuracy)	45.5%	37.6%	39.0%	47.0%
NLG	XLSum (ROUGE-L)	0.0452	0.0260	0.0277	0.0448
NLG	Flores200 (BLEU, en↔th avg)	34.1	37.0	33.3	38.3
NLG	iApp Wiki QA (F1)	67.6%	69.0%	51.2%	50.0%
Overall	normalized average²	0.444	0.405	0.305	0.349

The fine-tune outperforms its own base model on every dataset in both reasoning modes. Reasoning drives the largest gains on the Exam category specifically (ThaiExam +24.8pp, M3Exam +9.9pp think vs. no-think) with comparatively little effect on NLG quality — consistent with chain-of-thought helping multiple-choice/classification tasks more than open-ended generation.

¹ Qwen3.5-9B base was evaluated with reasoning forced off (enable_thinking: false). Left at its own default (reasoning on), it deliberates far past what's practical for a full run — e.g. looping through repeated draft translations of a single sentence and hitting a 2,048-token generation cap still undecided, where this fine-tune closes the same task in under 750 tokens. That gap — reasoning efficiently vs. not at all — is itself an effect of the fine-tuning.

² Mean of accuracy (Exam, NLU) and ROUGE-L / (BLEU÷100) / F1 (NLG). MT-Bench (LLM-as-judge) was not run — no judge model was configured for this evaluation pass.

Limitations

• Evaluation to date covers its own base model plus one other organization's model (Typhoon2.1-Gemma3-12B; see Evaluation) — no broader sweep across other Thai LLMs yet, and no LLM-as-judge (MT-Bench) pass.
• LoRA fine-tuning at this data scale (5M samples, ~5–10B+ tokens) may underfit relative to full fine-tuning; this is a known open question for the chosen rank and has not been directly measured against a full-FT counterpart.
• No dedicated preference-alignment (DPO/RLHF) pass yet — this is an SFT-only checkpoint. Response style, refusal calibration, and Thai-specific fluency/register are expected to improve in a planned follow-up stage.
• Domain coverage is uneven by language (see Training Data) — several domains exist in only one of the two languages, so cross-lingual transfer for those domains is untested.
• Not intended for safety-critical deployment without additional evaluation.

Contributor

• Atikan Payungthong (atikan@cmkl.ac.th)
• Dr. Akkarit Sangpetch (akkarit@cmkl.ac.th)

Related Artifacts

• Previous release: CMKL/MANGO-Qwen3-Omni-30B-A3B-Instruct
• Base model: Qwen/Qwen3.5-9B