---
base_model: google/gemma-4-31B-it
library_name: peft
model_name: gemma4-31b-it-glimmer-rp-r16a32
tags:
- base_model:adapter:google/gemma-4-31B-it
- lora
- sft
- transformers
- trl
licence: license
pipeline_tag: text-generation
---

# gemma4-31b-it-glimmer-rp-r16a32

This model is a fine-tuned version of [google/gemma-4-31B-it](https://huggingface.co/google/gemma-4-31B-it).

**W&B run:** [https://wandb.ai/cooawoo-personal/Gemma4-31B/runs/nbmb3v4h](https://wandb.ai/cooawoo-personal/Gemma4-31B/runs/nbmb3v4h)
## Training procedure

### Hyperparameters

| Parameter | Value |
|-----------|-------|
| Learning rate | `1e-05` |
| LR scheduler | rex (custom; max_lr=1e-5, min_lr=1e-6, warmup_ratio=0.05)_WITH_WARMUP |
| Per-device batch size | 1 |
| Gradient accumulation | 4 |
| Effective batch size | 4 |
| Epochs | 1 |
| Max sequence length | 6144 |
| Optimizer | OptimizerNames.PAGED_ADAMW_8BIT |
| Warmup ratio | 0.05 |
| Max gradient norm | 1.0 |
| Precision | bf16 |
| Loss type | nll |
| Assistant-only loss | yes |
| Chunked cross-entropy | yes |


### LoRA configuration

| Parameter | Value |
|-----------|-------|
| Rank (r) | 16 |
| Alpha | 32 |
| Target modules | .*language_model\.layers\.\d+\.(self_attn\.(q|k|v|o)_proj|mlp\.(gate|up|down)_proj)$ |
| Quantization | 4-bit (nf4) |


### Dataset statistics

| Dataset | Samples | Total tokens | Trainable tokens |
|---------|--------:|-------------:|-----------------:|
| writing_critique.jsonl | 1,586 | 1,317,233 | 599,216 |
| instruct.jsonl | 962 | 933,867 | 838,340 |
| marvin_style_bible.jsonl | 2,549 | 11,096,548 | 10,492,459 |
| rp_generation_mistral.jsonl | 255 | 825,572 | 375,698 |
| rp_analysis.jsonl | 244 | 725,115 | 177,861 |
| rp_generation_final.jsonl | 129 | 513,974 | 240,019 |
| **Total** | **5,725** | **15,412,309** | **12,723,593** |


<details>
<summary>Training config</summary>

```yaml
model_name_or_path: google/gemma-4-31B-it
data_config: data.yaml
prepared_dataset: prepared_packed
output_dir: gemma4-31b-it-glimmer-rp-r16a32
chat_template_path: chat_template_with_channel.jinja
attn_implementation: flex_attention
bf16: true
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
use_cce: true
chunked_mlp: true
chunked_mlp_chunks: 16
dataloader_num_workers: 2
dataloader_pin_memory: true
model_parallel: true
max_memory:
  0: 16GiB
  1: 24GiB
max_length: 6144
per_device_train_batch_size: 1
gradient_accumulation_steps: 4
pad_to_multiple_of: 128
use_peft: true
load_in_4bit: true
bnb_4bit_quant_type: nf4
lora_r: 16
lora_alpha: 32
lora_dropout: 0.0
use_rslora: false
lora_target_modules: .*language_model\.layers\.\d+\.(self_attn\.(q|k|v|o)_proj|mlp\.(gate|up|down)_proj)$
learning_rate: 1.0e-05
lr_scheduler_type: constant_with_warmup
warmup_ratio: 0.05
weight_decay: 0.0
max_grad_norm: 1.0
optim: paged_adamw_8bit
num_train_epochs: 1
saves_per_epoch: 2
save_total_limit: 4
rolling_save_steps: 30
rolling_save_total_limit: 1
assistant_only_loss: true
full_mask_reasoning: true
logging_steps: 1
disable_tqdm: false
report_to: wandb
run_name: g4-31b-it-glimmer-rp
```

</details>

<details>
<summary>Data config</summary>

```yaml
datasets:
- path: rp_generation_final.jsonl
  type: chat
  truncation_strategy: split
- path: rp_generation_mistral.jsonl
  type: chat
  truncation_strategy: split
- path: instruct.jsonl
  type: chat
  truncation_strategy: split
- path: rp_analysis.jsonl
  type: chat
  truncation_strategy: split
- path: writing_critique.jsonl
  type: chat
  truncation_strategy: split
- path: marvin_style_bible.jsonl
  type: chat
  truncation_strategy: split
shuffle_datasets: true
shuffle_combined: true
shuffle_seed: 42
eval_split: 0
split_seed: 42
assistant_only_loss: true
```

</details>

### Framework versions

- PEFT 0.18.1
- Loft: 0.1.0
- Transformers: 5.5.4
- Pytorch: 2.6.0+cu124
- Datasets: 4.6.1
- Tokenizers: 0.22.2