Gemma4-Gutenberg-12B
Training Configuration
| Parameter | Value |
|---|---|
| Training Mode | ORPO |
| Base Model | google/gemma-4-12B-it |
| Learning Rate | 5e-05 |
| Epochs | 1 |
| Batch Size | 1 |
| Gradient Accumulation | 32 |
| Effective Batch Size | 32 |
| Max Sequence Length | 1536 |
| Optimizer | paged_adamw_8bit |
| LR Scheduler | cosine |
| Warmup Ratio | 0.05 |
| Weight Decay | 0.01 |
| Max Grad Norm | 0.5 |
| Seed | 42 |
| Beta | 0.1 |
| Max Prompt Length | 1024 |
| LoRA Rank (r) | 64 |
| LoRA Alpha | 32 |
| LoRA Dropout | 0.05 |
| Target Modules | language_model.layers.0.self_attn.q_proj, language_model.layers.0.self_attn.k_proj, language_model.layers.0.self_attn.v_proj, language_model.layers.0.self_attn.o_proj, language_model.layers.0.mlp.gate_proj, language_model.layers.0.mlp.up_proj, language_model.layers.0.mlp.down_proj, language_model.layers.1.self_attn.q_proj, language_model.layers.1.self_attn.k_proj, language_model.layers.1.self_attn.v_proj, language_model.layers.1.self_attn.o_proj, language_model.layers.1.mlp.gate_proj, language_model.layers.1.mlp.up_proj, language_model.layers.1.mlp.down_proj, language_model.layers.2.self_attn.q_proj, language_model.layers.2.self_attn.k_proj, language_model.layers.2.self_attn.v_proj, language_model.layers.2.self_attn.o_proj, language_model.layers.2.mlp.gate_proj, language_model.layers.2.mlp.up_proj, language_model.layers.2.mlp.down_proj, language_model.layers.3.self_attn.q_proj, language_model.layers.3.self_attn.k_proj, language_model.layers.3.self_attn.v_proj, language_model.layers.3.self_attn.o_proj, language_model.layers.3.mlp.gate_proj, language_model.layers.3.mlp.up_proj, language_model.layers.3.mlp.down_proj, language_model.layers.4.self_attn.q_proj, language_model.layers.4.self_attn.k_proj, language_model.layers.4.self_attn.v_proj, language_model.layers.4.self_attn.o_proj, language_model.layers.4.mlp.gate_proj, language_model.layers.4.mlp.up_proj, language_model.layers.4.mlp.down_proj, language_model.layers.5.self_attn.q_proj, language_model.layers.5.self_attn.k_proj, language_model.layers.5.self_attn.v_proj, language_model.layers.5.self_attn.o_proj, language_model.layers.5.mlp.gate_proj, language_model.layers.5.mlp.up_proj, language_model.layers.5.mlp.down_proj, language_model.layers.6.self_attn.q_proj, language_model.layers.6.self_attn.k_proj, language_model.layers.6.self_attn.v_proj, language_model.layers.6.self_attn.o_proj, language_model.layers.6.mlp.gate_proj, language_model.layers.6.mlp.up_proj, language_model.layers.6.mlp.down_proj, language_model.layers.7.self_attn.q_proj, language_model.layers.7.self_attn.k_proj, language_model.layers.7.self_attn.v_proj, language_model.layers.7.self_attn.o_proj, language_model.layers.7.mlp.gate_proj, language_model.layers.7.mlp.up_proj, language_model.layers.7.mlp.down_proj, language_model.layers.8.self_attn.q_proj, language_model.layers.8.self_attn.k_proj, language_model.layers.8.self_attn.v_proj, language_model.layers.8.self_attn.o_proj, language_model.layers.8.mlp.gate_proj, language_model.layers.8.mlp.up_proj, language_model.layers.8.mlp.down_proj, language_model.layers.9.self_attn.q_proj, language_model.layers.9.self_attn.k_proj, language_model.layers.9.self_attn.v_proj, language_model.layers.9.self_attn.o_proj, language_model.layers.9.mlp.gate_proj, language_model.layers.9.mlp.up_proj, language_model.layers.9.mlp.down_proj, language_model.layers.10.self_attn.q_proj, language_model.layers.10.self_attn.k_proj, language_model.layers.10.self_attn.v_proj, language_model.layers.10.self_attn.o_proj, language_model.layers.10.mlp.gate_proj, language_model.layers.10.mlp.up_proj, language_model.layers.10.mlp.down_proj, language_model.layers.11.self_attn.q_proj, language_model.layers.11.self_attn.k_proj, language_model.layers.11.self_attn.v_proj, language_model.layers.11.self_attn.o_proj, language_model.layers.11.mlp.gate_proj, language_model.layers.11.mlp.up_proj, language_model.layers.11.mlp.down_proj, language_model.layers.12.self_attn.q_proj, language_model.layers.12.self_attn.k_proj, language_model.layers.12.self_attn.v_proj, language_model.layers.12.self_attn.o_proj, language_model.layers.12.mlp.gate_proj, language_model.layers.12.mlp.up_proj, language_model.layers.12.mlp.down_proj, language_model.layers.13.self_attn.q_proj, language_model.layers.13.self_attn.k_proj, language_model.layers.13.self_attn.v_proj, language_model.layers.13.self_attn.o_proj, language_model.layers.13.mlp.gate_proj, language_model.layers.13.mlp.up_proj, language_model.layers.13.mlp.down_proj, language_model.layers.14.self_attn.q_proj, language_model.layers.14.self_attn.k_proj, language_model.layers.14.self_attn.v_proj, language_model.layers.14.self_attn.o_proj, language_model.layers.14.mlp.gate_proj, language_model.layers.14.mlp.up_proj, language_model.layers.14.mlp.down_proj, language_model.layers.15.self_attn.q_proj, language_model.layers.15.self_attn.k_proj, language_model.layers.15.self_attn.v_proj, language_model.layers.15.self_attn.o_proj, language_model.layers.15.mlp.gate_proj, language_model.layers.15.mlp.up_proj, language_model.layers.15.mlp.down_proj, language_model.layers.16.self_attn.q_proj, language_model.layers.16.self_attn.k_proj, language_model.layers.16.self_attn.v_proj, language_model.layers.16.self_attn.o_proj, language_model.layers.16.mlp.gate_proj, language_model.layers.16.mlp.up_proj, language_model.layers.16.mlp.down_proj, language_model.layers.17.self_attn.q_proj, language_model.layers.17.self_attn.k_proj, language_model.layers.17.self_attn.v_proj, language_model.layers.17.self_attn.o_proj, language_model.layers.17.mlp.gate_proj, language_model.layers.17.mlp.up_proj, language_model.layers.17.mlp.down_proj, language_model.layers.18.self_attn.q_proj, language_model.layers.18.self_attn.k_proj, language_model.layers.18.self_attn.v_proj, language_model.layers.18.self_attn.o_proj, language_model.layers.18.mlp.gate_proj, language_model.layers.18.mlp.up_proj, language_model.layers.18.mlp.down_proj, language_model.layers.19.self_attn.q_proj, language_model.layers.19.self_attn.k_proj, language_model.layers.19.self_attn.v_proj, language_model.layers.19.self_attn.o_proj, language_model.layers.19.mlp.gate_proj, language_model.layers.19.mlp.up_proj, language_model.layers.19.mlp.down_proj, language_model.layers.20.self_attn.q_proj, language_model.layers.20.self_attn.k_proj, language_model.layers.20.self_attn.v_proj, language_model.layers.20.self_attn.o_proj, language_model.layers.20.mlp.gate_proj, language_model.layers.20.mlp.up_proj, language_model.layers.20.mlp.down_proj, language_model.layers.21.self_attn.q_proj, language_model.layers.21.self_attn.k_proj, language_model.layers.21.self_attn.v_proj, language_model.layers.21.self_attn.o_proj, language_model.layers.21.mlp.gate_proj, language_model.layers.21.mlp.up_proj, language_model.layers.21.mlp.down_proj, language_model.layers.22.self_attn.q_proj, language_model.layers.22.self_attn.k_proj, language_model.layers.22.self_attn.v_proj, language_model.layers.22.self_attn.o_proj, language_model.layers.22.mlp.gate_proj, language_model.layers.22.mlp.up_proj, language_model.layers.22.mlp.down_proj, language_model.layers.23.self_attn.q_proj, language_model.layers.23.self_attn.k_proj, language_model.layers.23.self_attn.v_proj, language_model.layers.23.self_attn.o_proj, language_model.layers.23.mlp.gate_proj, language_model.layers.23.mlp.up_proj, language_model.layers.23.mlp.down_proj, language_model.layers.24.self_attn.q_proj, language_model.layers.24.self_attn.k_proj, language_model.layers.24.self_attn.v_proj, language_model.layers.24.self_attn.o_proj, language_model.layers.24.mlp.gate_proj, language_model.layers.24.mlp.up_proj, language_model.layers.24.mlp.down_proj, language_model.layers.25.self_attn.q_proj, language_model.layers.25.self_attn.k_proj, language_model.layers.25.self_attn.v_proj, language_model.layers.25.self_attn.o_proj, language_model.layers.25.mlp.gate_proj, language_model.layers.25.mlp.up_proj, language_model.layers.25.mlp.down_proj, language_model.layers.26.self_attn.q_proj, language_model.layers.26.self_attn.k_proj, language_model.layers.26.self_attn.v_proj, language_model.layers.26.self_attn.o_proj, language_model.layers.26.mlp.gate_proj, language_model.layers.26.mlp.up_proj, language_model.layers.26.mlp.down_proj, language_model.layers.27.self_attn.q_proj, language_model.layers.27.self_attn.k_proj, language_model.layers.27.self_attn.v_proj, language_model.layers.27.self_attn.o_proj, language_model.layers.27.mlp.gate_proj, language_model.layers.27.mlp.up_proj, language_model.layers.27.mlp.down_proj, language_model.layers.28.self_attn.q_proj, language_model.layers.28.self_attn.k_proj, language_model.layers.28.self_attn.v_proj, language_model.layers.28.self_attn.o_proj, language_model.layers.28.mlp.gate_proj, language_model.layers.28.mlp.up_proj, language_model.layers.28.mlp.down_proj, language_model.layers.29.self_attn.q_proj, language_model.layers.29.self_attn.k_proj, language_model.layers.29.self_attn.v_proj, language_model.layers.29.self_attn.o_proj, language_model.layers.29.mlp.gate_proj, language_model.layers.29.mlp.up_proj, language_model.layers.29.mlp.down_proj |
| Quantization | 4-bit (NF4) |
| GPU | NVIDIA GB10 |
Reproduce this training run
This model was trained with Merlina. Save the JSON below to data/configs/<name>.json (or import it via the Load Configuration dialog) to reproduce the exact training setup. Credentials are not included — Merlina will use your own HF_TOKEN and WANDB_API_KEY from .env or the form.
{
"_metadata": {
"name": "Gemma4-Gutenberg-12B",
"description": "Training configuration shared from a Merlina-trained model.",
"tags": [],
"schema": "merlina/training-config",
"schema_version": 1,
"merlina_version": "2.0.3"
},
"base_model": "google/gemma-4-12B-it",
"output_name": "Gemma4-Gutenberg-12B",
"use_lora": true,
"lora_r": 64,
"lora_alpha": 32,
"lora_dropout": 0.05,
"target_modules": [
"language_model.layers.0.self_attn.q_proj",
"language_model.layers.0.self_attn.k_proj",
"language_model.layers.0.self_attn.v_proj",
"language_model.layers.0.self_attn.o_proj",
"language_model.layers.0.mlp.gate_proj",
"language_model.layers.0.mlp.up_proj",
"language_model.layers.0.mlp.down_proj",
"language_model.layers.1.self_attn.q_proj",
"language_model.layers.1.self_attn.k_proj",
"language_model.layers.1.self_attn.v_proj",
"language_model.layers.1.self_attn.o_proj",
"language_model.layers.1.mlp.gate_proj",
"language_model.layers.1.mlp.up_proj",
"language_model.layers.1.mlp.down_proj",
"language_model.layers.2.self_attn.q_proj",
"language_model.layers.2.self_attn.k_proj",
"language_model.layers.2.self_attn.v_proj",
"language_model.layers.2.self_attn.o_proj",
"language_model.layers.2.mlp.gate_proj",
"language_model.layers.2.mlp.up_proj",
"language_model.layers.2.mlp.down_proj",
"language_model.layers.3.self_attn.q_proj",
"language_model.layers.3.self_attn.k_proj",
"language_model.layers.3.self_attn.v_proj",
"language_model.layers.3.self_attn.o_proj",
"language_model.layers.3.mlp.gate_proj",
"language_model.layers.3.mlp.up_proj",
"language_model.layers.3.mlp.down_proj",
"language_model.layers.4.self_attn.q_proj",
"language_model.layers.4.self_attn.k_proj",
"language_model.layers.4.self_attn.v_proj",
"language_model.layers.4.self_attn.o_proj",
"language_model.layers.4.mlp.gate_proj",
"language_model.layers.4.mlp.up_proj",
"language_model.layers.4.mlp.down_proj",
"language_model.layers.5.self_attn.q_proj",
"language_model.layers.5.self_attn.k_proj",
"language_model.layers.5.self_attn.v_proj",
"language_model.layers.5.self_attn.o_proj",
"language_model.layers.5.mlp.gate_proj",
"language_model.layers.5.mlp.up_proj",
"language_model.layers.5.mlp.down_proj",
"language_model.layers.6.self_attn.q_proj",
"language_model.layers.6.self_attn.k_proj",
"language_model.layers.6.self_attn.v_proj",
"language_model.layers.6.self_attn.o_proj",
"language_model.layers.6.mlp.gate_proj",
"language_model.layers.6.mlp.up_proj",
"language_model.layers.6.mlp.down_proj",
"language_model.layers.7.self_attn.q_proj",
"language_model.layers.7.self_attn.k_proj",
"language_model.layers.7.self_attn.v_proj",
"language_model.layers.7.self_attn.o_proj",
"language_model.layers.7.mlp.gate_proj",
"language_model.layers.7.mlp.up_proj",
"language_model.layers.7.mlp.down_proj",
"language_model.layers.8.self_attn.q_proj",
"language_model.layers.8.self_attn.k_proj",
"language_model.layers.8.self_attn.v_proj",
"language_model.layers.8.self_attn.o_proj",
"language_model.layers.8.mlp.gate_proj",
"language_model.layers.8.mlp.up_proj",
"language_model.layers.8.mlp.down_proj",
"language_model.layers.9.self_attn.q_proj",
"language_model.layers.9.self_attn.k_proj",
"language_model.layers.9.self_attn.v_proj",
"language_model.layers.9.self_attn.o_proj",
"language_model.layers.9.mlp.gate_proj",
"language_model.layers.9.mlp.up_proj",
"language_model.layers.9.mlp.down_proj",
"language_model.layers.10.self_attn.q_proj",
"language_model.layers.10.self_attn.k_proj",
"language_model.layers.10.self_attn.v_proj",
"language_model.layers.10.self_attn.o_proj",
"language_model.layers.10.mlp.gate_proj",
"language_model.layers.10.mlp.up_proj",
"language_model.layers.10.mlp.down_proj",
"language_model.layers.11.self_attn.q_proj",
"language_model.layers.11.self_attn.k_proj",
"language_model.layers.11.self_attn.v_proj",
"language_model.layers.11.self_attn.o_proj",
"language_model.layers.11.mlp.gate_proj",
"language_model.layers.11.mlp.up_proj",
"language_model.layers.11.mlp.down_proj",
"language_model.layers.12.self_attn.q_proj",
"language_model.layers.12.self_attn.k_proj",
"language_model.layers.12.self_attn.v_proj",
"language_model.layers.12.self_attn.o_proj",
"language_model.layers.12.mlp.gate_proj",
"language_model.layers.12.mlp.up_proj",
"language_model.layers.12.mlp.down_proj",
"language_model.layers.13.self_attn.q_proj",
"language_model.layers.13.self_attn.k_proj",
"language_model.layers.13.self_attn.v_proj",
"language_model.layers.13.self_attn.o_proj",
"language_model.layers.13.mlp.gate_proj",
"language_model.layers.13.mlp.up_proj",
"language_model.layers.13.mlp.down_proj",
"language_model.layers.14.self_attn.q_proj",
"language_model.layers.14.self_attn.k_proj",
"language_model.layers.14.self_attn.v_proj",
"language_model.layers.14.self_attn.o_proj",
"language_model.layers.14.mlp.gate_proj",
"language_model.layers.14.mlp.up_proj",
"language_model.layers.14.mlp.down_proj",
"language_model.layers.15.self_attn.q_proj",
"language_model.layers.15.self_attn.k_proj",
"language_model.layers.15.self_attn.v_proj",
"language_model.layers.15.self_attn.o_proj",
"language_model.layers.15.mlp.gate_proj",
"language_model.layers.15.mlp.up_proj",
"language_model.layers.15.mlp.down_proj",
"language_model.layers.16.self_attn.q_proj",
"language_model.layers.16.self_attn.k_proj",
"language_model.layers.16.self_attn.v_proj",
"language_model.layers.16.self_attn.o_proj",
"language_model.layers.16.mlp.gate_proj",
"language_model.layers.16.mlp.up_proj",
"language_model.layers.16.mlp.down_proj",
"language_model.layers.17.self_attn.q_proj",
"language_model.layers.17.self_attn.k_proj",
"language_model.layers.17.self_attn.v_proj",
"language_model.layers.17.self_attn.o_proj",
"language_model.layers.17.mlp.gate_proj",
"language_model.layers.17.mlp.up_proj",
"language_model.layers.17.mlp.down_proj",
"language_model.layers.18.self_attn.q_proj",
"language_model.layers.18.self_attn.k_proj",
"language_model.layers.18.self_attn.v_proj",
"language_model.layers.18.self_attn.o_proj",
"language_model.layers.18.mlp.gate_proj",
"language_model.layers.18.mlp.up_proj",
"language_model.layers.18.mlp.down_proj",
"language_model.layers.19.self_attn.q_proj",
"language_model.layers.19.self_attn.k_proj",
"language_model.layers.19.self_attn.v_proj",
"language_model.layers.19.self_attn.o_proj",
"language_model.layers.19.mlp.gate_proj",
"language_model.layers.19.mlp.up_proj",
"language_model.layers.19.mlp.down_proj",
"language_model.layers.20.self_attn.q_proj",
"language_model.layers.20.self_attn.k_proj",
"language_model.layers.20.self_attn.v_proj",
"language_model.layers.20.self_attn.o_proj",
"language_model.layers.20.mlp.gate_proj",
"language_model.layers.20.mlp.up_proj",
"language_model.layers.20.mlp.down_proj",
"language_model.layers.21.self_attn.q_proj",
"language_model.layers.21.self_attn.k_proj",
"language_model.layers.21.self_attn.v_proj",
"language_model.layers.21.self_attn.o_proj",
"language_model.layers.21.mlp.gate_proj",
"language_model.layers.21.mlp.up_proj",
"language_model.layers.21.mlp.down_proj",
"language_model.layers.22.self_attn.q_proj",
"language_model.layers.22.self_attn.k_proj",
"language_model.layers.22.self_attn.v_proj",
"language_model.layers.22.self_attn.o_proj",
"language_model.layers.22.mlp.gate_proj",
"language_model.layers.22.mlp.up_proj",
"language_model.layers.22.mlp.down_proj",
"language_model.layers.23.self_attn.q_proj",
"language_model.layers.23.self_attn.k_proj",
"language_model.layers.23.self_attn.v_proj",
"language_model.layers.23.self_attn.o_proj",
"language_model.layers.23.mlp.gate_proj",
"language_model.layers.23.mlp.up_proj",
"language_model.layers.23.mlp.down_proj",
"language_model.layers.24.self_attn.q_proj",
"language_model.layers.24.self_attn.k_proj",
"language_model.layers.24.self_attn.v_proj",
"language_model.layers.24.self_attn.o_proj",
"language_model.layers.24.mlp.gate_proj",
"language_model.layers.24.mlp.up_proj",
"language_model.layers.24.mlp.down_proj",
"language_model.layers.25.self_attn.q_proj",
"language_model.layers.25.self_attn.k_proj",
"language_model.layers.25.self_attn.v_proj",
"language_model.layers.25.self_attn.o_proj",
"language_model.layers.25.mlp.gate_proj",
"language_model.layers.25.mlp.up_proj",
"language_model.layers.25.mlp.down_proj",
"language_model.layers.26.self_attn.q_proj",
"language_model.layers.26.self_attn.k_proj",
"language_model.layers.26.self_attn.v_proj",
"language_model.layers.26.self_attn.o_proj",
"language_model.layers.26.mlp.gate_proj",
"language_model.layers.26.mlp.up_proj",
"language_model.layers.26.mlp.down_proj",
"language_model.layers.27.self_attn.q_proj",
"language_model.layers.27.self_attn.k_proj",
"language_model.layers.27.self_attn.v_proj",
"language_model.layers.27.self_attn.o_proj",
"language_model.layers.27.mlp.gate_proj",
"language_model.layers.27.mlp.up_proj",
"language_model.layers.27.mlp.down_proj",
"language_model.layers.28.self_attn.q_proj",
"language_model.layers.28.self_attn.k_proj",
"language_model.layers.28.self_attn.v_proj",
"language_model.layers.28.self_attn.o_proj",
"language_model.layers.28.mlp.gate_proj",
"language_model.layers.28.mlp.up_proj",
"language_model.layers.28.mlp.down_proj",
"language_model.layers.29.self_attn.q_proj",
"language_model.layers.29.self_attn.k_proj",
"language_model.layers.29.self_attn.v_proj",
"language_model.layers.29.self_attn.o_proj",
"language_model.layers.29.mlp.gate_proj",
"language_model.layers.29.mlp.up_proj",
"language_model.layers.29.mlp.down_proj"
],
"modules_to_save": [],
"lora_task_type": "CAUSAL_LM",
"learning_rate": 5e-05,
"num_epochs": 1,
"batch_size": 1,
"gradient_accumulation_steps": 32,
"max_length": 1536,
"max_prompt_length": 1024,
"model_type": "auto",
"training_mode": "orpo",
"beta": 0.1,
"label_smoothing": 0.0,
"gamma": 0.5,
"vision_model_id": null,
"stage": null,
"unfreeze_vision_top_n": null,
"image_token_id": null,
"min_pixels": null,
"max_pixels": null,
"image_column": null,
"caption_column": null,
"instruction": null,
"streaming": null,
"model_name": null,
"image_resolution": 1024,
"lora_rank": 32,
"lora_target_modules": null,
"lora_use_dora": false,
"mid_training_samples": true,
"dataset_jsonl_path": null,
"dataset_name": null,
"dataset_split": null,
"sample_prompts": null,
"sample_num_steps": null,
"dataset": {
"source": {
"source_type": "huggingface",
"repo_id": "schneewolflabs/Alembic-DPO",
"split": "train",
"file_path": null,
"file_format": null,
"dataset_id": null,
"streaming": false,
"streaming_batch_size": 10000,
"column_mapping": null
},
"additional_sources": [],
"format": {
"format_type": "tokenizer",
"custom_templates": null,
"enable_thinking": true,
"auto_detect_thinking": true
},
"model_name": "google/gemma-4-12B-it",
"column_mapping": {
"prompt": "prompt",
"chosen": "chosen",
"rejected": "rejected",
"system": "system"
},
"convert_messages_format": true,
"deduplicate": false,
"dedupe_strategy": "prompt_chosen",
"test_size": 0.02,
"max_samples": null,
"system_prompt": null,
"system_prompt_mode": "fill_empty",
"training_mode": "orpo"
},
"seed": 42,
"max_grad_norm": 0.5,
"warmup_ratio": 0.05,
"eval_steps": 0.2,
"use_4bit": true,
"use_wandb": true,
"push_to_hub": false,
"merge_lora_before_upload": true,
"hf_hub_private": false,
"export_gguf": false,
"gguf_quant_types": [
"Q4_K_M"
],
"keep_gguf_fp16": false,
"shuffle_dataset": true,
"weight_decay": 0.01,
"lr_scheduler_type": "cosine",
"gradient_checkpointing": true,
"logging_steps": 1,
"optimizer_type": "paged_adamw_8bit",
"adam_beta1": 0.9,
"adam_beta2": 0.999,
"adam_epsilon": 1e-08,
"adafactor_relative_step": false,
"adafactor_scale_parameter": false,
"adafactor_warmup_init": false,
"adafactor_decay_rate": -0.8,
"adafactor_beta1": null,
"adafactor_clip_threshold": 1.0,
"attn_implementation": "auto",
"use_liger": false,
"torch_compile": false,
"neftune_alpha": null,
"eval_on_start": false,
"gpu_ids": null,
"multi_gpu_strategy": "auto",
"wandb_project": "gutenberg",
"wandb_run_name": null,
"wandb_tags": null,
"wandb_notes": null
}
