GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
IRedDragonICY
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
IRedDragonICY
Model Tree
openbmb/MiniCPM5-2B
Input Modalities
Output Modalities
Supported Functionality
enable_thinking=FalseThe MiniCPM5 chat template inserts an empty <think></think> block in front
of every assistant turn in the training data, but adds nothing after
<|im_start|>assistant when enable_thinking is left undefined. The adapter
learned to answer after the empty block, so prompts must be built with
enable_thinking=False. The table below shows how much this matters.
Accuracy on the 1,200 LaguQA multiple-choice questions, scored from the mean log-probability of each option's text. One seed.
| Prompt template | Before training | After training | Change |
|---|---|---|---|
| Template default (protocol used for all 14 open-weight models) | 24.2 | 36.9 | +12.8 |
enable_thinking=False (same format as the training data) | 24.3 | 51.0 | +26.8 |
Per question group, enable_thinking=False:
| Overall | Facts | Title | Lyrics | Notation | Reasoning | |
|---|---|---|---|---|---|---|
| Before training | 24.3 | 18.5 | 23.4 | 22.3 | 22.3 | 34.8 |
| After training | 51.0 | 52.4 | 16.8 | 26.0 | 61.1 | 75.9 |
| Answer-distribution baseline | 32.1 | 58.4 | 17.8 | 24.0 | 17.6 | 20.3 |
Under the template default, 192 questions went from wrong to right and 39 from right to wrong (McNemar p ≈ 1.8 × 10⁻²⁵). For comparison, the Gemma 4 E2B adapter trained with the same recipe rose from 24.1 to 53.3 (mean of three seeds).
IndoMMLU, 2,000 questions, neutral system prompt, random guessing 23.4%.
| Prompt template | Before training | After training | Change |
|---|---|---|---|
| Template default | 28.4 | 25.8 | −2.6 |
enable_thinking=False | 33.6 | 27.2 | −6.4 |
Unlike Gemma 4 E2B, whose IndoMMLU score rose after the same training, this model loses general Indonesian knowledge. Use the base model for anything outside the songbook.
python
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer dasar = "openbmb/MiniCPM5-2B"tok = AutoTokenizer.from_pretrained(dasar)model = AutoModelForCausalLM.from_pretrained(dasar, dtype="bfloat16")model = PeftModel.from_pretrained(model, "IRedDragonICY/LaguQA-MiniCPM5-2B-LoRA") pesan = [{"role": "system", "content": "Kamu asisten yang menguasai lagu nasional dan lagu daerah " "Indonesia, termasuk notasi angka dan notasi ABC-nya."}, {"role": "user", "content": "Siapa pencipta lagu Gugur Bunga?"}]teks = tok.apply_chat_template(pesan, tokenize=False, add_generation_prompt=True, enable_thinking=False)keluar = model.generate(**tok(teks, return_tensors="pt").to(model.device), max_new_tokens=192, do_sample=False)print(tok.decode(keluar[0], skip_special_tokens=True))| Training file | laguqa_train.jsonl, sha256 70d497ae… (20,944 training and 1,102 validation examples, split with the run seed) |
| LoRA | r 16, alpha 32, dropout 0.05 on q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj; 294 modules in 42 layers |
| Trained parameters | 25,116,672 of 2,541,873,152 (0.99%) |
| Schedule | 2 epochs, 2,618 steps, learning rate 4e-4, cosine, 78 warmup steps, batch 4 with 4 accumulation steps, loss on the assistant turn only, sequences cut at 768 tokens, seed 1 |
| Hardware | one NVIDIA Tesla T4 on Kaggle Notebooks, float16 with a gradient scaler (the T4 has no bfloat16), 3.8 s per step, 2.8 hours |
| Loss | training 3.196 → 0.083; validation 0.244 (epoch 1) → 0.110 (epoch 2) |
| Libraries | torch 2.10.0, transformers 5.16.1, peft 0.20.0, trl 1.12.0 |
The recipe is identical to the Gemma 4 E2B "final" variant except for the
precision, which the T4 forced. Full settings are in manifest.json.
Adapter weights: Apache-2.0, the licence of openbmb/MiniCPM5-2B. Data: CC BY-NC
4.0; per-song rights are listed in HAK-CIPTA.md in the dataset repository.
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer dasar = "openbmb/MiniCPM5-2B"tok = AutoTokenizer.from_pretrained(dasar)model = AutoModelForCausalLM.from_pretrained(dasar, dtype="bfloat16")model = PeftModel.from_pretrained(model, "IRedDragonICY/LaguQA-MiniCPM5-2B-LoRA") pesan = [{"role": "system", "content": "Kamu asisten yang menguasai lagu nasional dan lagu daerah " "Indonesia, termasuk notasi angka dan notasi ABC-nya."}, {"role": "user", "content": "Siapa pencipta lagu Gugur Bunga?"}]teks = tok.apply_chat_template(pesan, tokenize=False, add_generation_prompt=True, enable_thinking=False)keluar = model.generate(**tok(teks, return_tensors="pt").to(model.device), max_new_tokens=192, do_sample=False)print(tok.decode(keluar[0], skip_special_tokens=True))