GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
GL3MON
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
GL3MON
Model Tree
Qwen/Qwen3.5-2B
Input Modalities
Output Modalities
Supported Functionality
Base: Qwen/Qwen3.5-2B. A fresh text-only LoRA (rank 16, alpha 32) was supervised on 2,000 Tamil GSM8K TRAIN questions paired with faithful original English questions and official worked solutions, for one epoch / 500 optimizer steps. Completion-only loss masks prompt tokens. A disjoint 128-question TRAIN validation subset selected the SFT checkpoint and was excluded from the subsequent RL pool.
RL initializes the same adapter from SFT and uses a local TAPO adaptation on TRL GRPO: separate group-normalized translation and correctness rewards, token-dependent translation/reasoning advantages, alpha 0.25, beta 0, DAPO loss, and 8 completions per prompt. Translation reward is ChrF++ against the original English question; correctness requires the correct boxed answer after a valid translation block. This is the project's adaptation, not an official TAPO implementation. Configuration is included in training/tapo_config.yaml.
The RL dataset contains 7,345 GSM8K TRAIN questions translated into Tamil. Official GSM8K test questions were not used for SFT, RL, or SFT checkpoint selection. Original worked solutions originate from openai/gsm8k; Tamil training questions come from the project's translated GSM8K training file. Vishal0407/GSM8K_TAMIL was used for evaluation only.
Use the official Qwen chat template with hidden thinking disabled (enable_thinking=False). The expected response translates the question into English inside <english_translation>...</english_translation>, then reasons in English and ends with \boxed{...}.
python
import torchfrom transformers import AutoTokenizer, Qwen3_5ForConditionalGeneration model_id = "GL3MON/Qwen3.5-2B-Tamil-TAPO"tokenizer = AutoTokenizer.from_pretrained(model_id)model = Qwen3_5ForConditionalGeneration.from_pretrained( model_id, dtype=torch.bfloat16, device_map="auto")messages = [ {"role": "system", "content": 'Solve the mathematical word problem. First translate the complete question faithfully into English inside <english_translation>...</english_translation>. Preserve all quantities, units, and relationships. Then reason in English and give the final answer in \\boxed{...}. Do not put reasoning inside the translation tags.'}, {"role": "user", "content": "ஒரு பெட்டியில் 12 பந்துகள் உள்ளன. மேலும் 8 பந்துகள் சேர்த்தால் மொத்தம் எத்தனை?"},]inputs = tokenizer.apply_chat_template( messages, tokenize=True, return_dict=True, return_tensors="pt", add_generation_prompt=True, enable_thinking=False,).to(model.device)with torch.inference_mode(): output = model.generate(**inputs, do_sample=False, max_new_tokens=1536)print(tokenizer.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))Tested training/inference environment: PyTorch 2.11, Transformers 5.6.2, PEFT 0.20, TRL 1.9.2, and vLLM 0.22.0. Text inference is the intended use; only language-model LoRA targets were trained.
This uploaded checkpoint (1525) has not yet been evaluated. The following results describe earlier checkpoints, not this upload:
| Checkpoint | Correct on fixed 500-question Tamil subset | Accuracy |
|---|---|---|
| 2,000-example SFT | 180/500 | 36.0% |
| RL step 1125 | 184/500 | 36.8% |
| RL step 1375 | 186/500 | 37.2% |
The subset uses seed 42 and 500 questions from the inferred official-test portion of Vishal0407/GSM8K_TAMIL, whose combined train/test ordering was checked against official final answers and calculation signatures. Evaluation uses one greedy answer per question and a 1,536-token completion cap. These are not full English GSM8K scores, Coverage@8, or combined retry scores. Translation quality can affect results, and the observed gains are small. Earlier checkpoint evaluation details are included in earlier_checkpoint_evaluation.json.
Apache-2.0, following the base model's license. See the original Qwen3.5-2B model card.
import torchfrom transformers import AutoTokenizer, Qwen3_5ForConditionalGeneration model_id = "GL3MON/Qwen3.5-2B-Tamil-TAPO"tokenizer = AutoTokenizer.from_pretrained(model_id)model = Qwen3_5ForConditionalGeneration.from_pretrained( model_id, dtype=torch.bfloat16, device_map="auto")messages = [ {"role": "system", "content": 'Solve the mathematical word problem. First translate the complete question faithfully into English inside <english_translation>...</english_translation>. Preserve all quantities, units, and relationships. Then reason in English and give the final answer in \\boxed{...}. Do not put reasoning inside the translation tags.'}, {"role": "user", "content": "ஒரு பெட்டியில் 12 பந்துகள் உள்ளன. மேலும் 8 பந்துகள் சேர்த்தால் மொத்தம் எத்தனை?"},]inputs = tokenizer.apply_chat_template( messages, tokenize=True, return_dict=True, return_tensors="pt", add_generation_prompt=True, enable_thinking=False,).to(model.device)with torch.inference_mode(): output = model.generate(**inputs, do_sample=False, max_new_tokens=1536)print(tokenizer.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))