Views
No views yet
1{
2 "model": "Qwen3-30B-A3B-Instruct-2507",
3 "phase": "phase2a-test-1k",
4 "dataset": "phase2-rp-base-1k",
5 "num_samples": 1000,
6 "lora_rank": 16,
7 "lora_alpha": 16,
8 "lora_dropout": 0,
9 "learning_rate": 0.0002,
10 "batch_size": 2,
11 "gradient_accumulation_steps": 4,
12 "effective_batch_size": 32,
13 "max_steps": 100,
14 "warmup_steps": 10,
15 "max_seq_length": 2048,
16 "optimizer": "adamw_8bit",
17 "weight_decay": 0.01,
18 "lr_scheduler_type": "linear",
19 "precision": "bfloat16",
20 "device_map": "auto",
21 "gpus": "4x RTX 5090",
22 "training_time": "40 minutes",
23 "framework": "Unsloth 2025.11.3",
24 "target_modules": [
25 "q_proj",
26 "k_proj",
27 "v_proj",
28 "o_proj",
29 "gate_proj",
30 "up_proj",
31 "down_proj"
32 ]
33}1{
2 "training_loss": {
3 "initial": 2.3745,
4 "final": 1.5027,
5 "reduction_percent": 36.7
6 },
7 "training_metrics": {
8 "total_steps": 100,
9 "total_samples": 1000,
10 "training_time_seconds": 2380.49,
11 "training_time_minutes": 39.67,
12 "samples_per_second": 0.336,
13 "final_grad_norm": 0.1539,
14 "final_learning_rate": 0.0
15 },
16 "loss_progression": {
17 "step_5": 2.3745,
18 "step_10": 1.531,
19 "step_50": 1.632,
20 "step_100": 1.5027
21 },
22 "wandb_run": "https://wandb.ai/developer_lunark-lunark-ai/kaidol-llm-finetuning/runs/brryct5m",
23 "notes": "Baseline test with 1K samples. Stable convergence observed. Ready for hyperparameter optimization (LR 2e-4→1e-4, alpha 16→32, grad_accum 4→8)."
24}1from unsloth import FastLanguageModel
2
3model, tokenizer = FastLanguageModel.from_pretrained(
4 model_name="developer-lunark/kaidol-phase2a-test-1k",
5 max_seq_length=2048,
6 dtype=None,
7 load_in_4bit=True,
8)1messages = [
2 {"role": "user", "content": "오늘 기분이 좋지 않아..."},
3]
4
5inputs = tokenizer.apply_chat_template(
6 messages,
7 tokenize=True,
8 add_generation_prompt=True,
9 return_tensors="pt"
10).to("cuda")
11
12outputs = model.generate(
13 inputs,
14 max_new_tokens=512,
15 temperature=0.7,
16 top_p=0.9,
17)
18
19response = tokenizer.decode(outputs[0], skip_special_tokens=True)
20print(response)developer-lunark/kaidol-phase2-rp-base-v0.11@misc{kaidol-phase2a-test-1k,
2 author = {Developer Lunark},
3 title = {KAIDOL LLM Fine-tuning - PHASE2A-TEST-1K},
4 year = {2025},
5 publisher = {HuggingFace},
6 howpublished = {\url{https://huggingface.co/developer-lunark/kaidol-phase2a-test-1k}}
7}