A LORA adapter for
google/gemma-4-31B-it-VLM. This model was trained with SFT using
Adaption's AutoScientist on the multilingual_vqa_test dataset.
1{
2 "job_id": "c41cc068-7da6-494c-820d-f2ee12be08e6",
3 "training_experiment_id": "6f6ddf31-899a-48f3-8bfd-abd7590148ab",
4 "original_model_name": "google/gemma-4-31B-it-VLM",
5 "trained_model_name": "adaption_multilingual_vqa_test",
6 "training_method": "sft",
7 "training_type": "lora",
8 "data_format": "chat",
9 "hyperparams": {
10 "lora": "true",
11 "lora_r": 8,
12 "n_evals": 5,
13 "n_epochs": 1,
14 "batch_size": "max",
15 "lora_alpha": 8,
16 "lora_dropout": 0,
17 "min_lr_ratio": 0.1,
18 "warmup_ratio": 0.1,
19 "weight_decay": 0,
20 "learning_rate": 0.00005,
21 "max_grad_norm": 2,
22 "base_model_size": "31B",
23 "train_on_inputs": "false",
24 "training_method": "sft",
25 "lr_scheduler_type": "cosine",
26 "scheduler_num_cycles": 0.5,
27 "lora_trainable_modules": "q_proj,v_proj"
28 }
29}
The model was trained on 819 rows of adapted data with the following domain distribution: math (32%), data-analysis-visualization (15%), science (6%), language (5%), other (5%), architecture-design (5%), corporate-business (5%), fitness-sports (4%), transportation (4%), animal-nature (2%), sports (2%), cooking (2%), geography (2%), art (2%), fashion-beauty (1%), academic-education (1%), culture (1%), product-advice (1%), market-analysis (1%), music (1%), personal-growth (1%), code (0%), entertainment (0%), technology (0%), marketing (0%), governance (0%).
The model was evaluated on an in-distribution held-out test set as well as a broader domain-specific test set to measure generalization.
1import torch
2from transformers import AutoModelForCausalLM, AutoTokenizer
3from peft import PeftModel
4
5BASE = "google/gemma-4-31B-it-VLM"
6ADAPTER = "<this-repo-id>"
7
8device = "cuda" if torch.cuda.is_available() else "cpu"
9dtype = torch.float32 if device == "cpu" else torch.bfloat16
10
11base = AutoModelForCausalLM.from_pretrained(BASE, dtype=dtype).to(device)
12model = PeftModel.from_pretrained(base, ADAPTER)
13# Optional: merge the LoRA weights into the base for faster inference
14model = model.merge_and_unload()
15model.eval()
16
17tokenizer = AutoTokenizer.from_pretrained(BASE)
18messages = [{"role": "user", "content": "Hello!"}]
19text = tokenizer.apply_chat_template(
20 messages, tokenize=False, add_generation_prompt=True)
21inputs = tokenizer(text, return_tensors="pt").to(device)
22
23with torch.inference_mode():
24 out = model.generate(**inputs, max_new_tokens=512)
25print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
