Downloads · 30 days
5
14% of all-time downloads
gemma4-31b-augmental-sft-ep1
nshuster/gemma4-31b-augmental-sft-ep1
gemma4-31b-augmental-sft-ep1 is a machine learning model from nshuster. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as gemma.
Full-parameter SFT of google/gemma-4-31B (base) on Heralax/Augmental-Dataset (7,831 rows, visual-novel style multi-character roleplay dialogue).
All-time downloads
37
Public
Parameters
1.5M
62.6 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors62.5 GB · 100%
At a glance
- License
- gemma
- Model type
- gemma4
- Access
- Public
- Created
- Jul 18, 2026
- Updated
- Jul 18, 2026
- SHA
- eac7c098
Base models
Datasets
Model card
From the Hugging Face model README
gemma4-31b-augmental-sft — epoch 1
Full-parameter SFT of google/gemma-4-31B (base) on Heralax/Augmental-Dataset (7,831 rows, visual-novel style multi-character roleplay dialogue).
This is the end-of-epoch-1 checkpoint (240/720 steps). Sibling repos: epoch 2 (-ep2), epoch 3 (-ep3).
Eval
| Checkpoint | eval_loss | eval_ppl |
|---|---|---|
| base (step 0) | 1.753 | 5.77 |
| epoch 1 (this repo) | 1.572 | 4.81 |
Prompt format
Trained with a plain-text scenario format (no chat template). Prompt the model exactly like this, then let it continue after the trailing speaker tag:
Scenario: {scenario description}
{Speaker A}: "…"
{Speaker B}: "…"
{target speaker}:
Generation ends with <eos>.
Training config
- 8× H200, DeepSpeed ZeRO-3, bf16, SDPA attention (FlashAttention-2 is incompatible with Gemma-4's 512-dim global-attention layers)
- seq len 4096, global batch 32 (micro 2 × accum 2 × 8 GPUs)
- lr 5e-6 cosine, warmup 25 steps, 3 epochs total, prompt tokens masked (
train_on_inputs: false) - Trained with axolotl
License
Gemma derivatives are governed by the Gemma Terms of Use, including the Gemma Prohibited Use Policy.
- Type
- gemma4
- License
- gemma
- Created
- Jul 18, 2026
- Updated
- Jul 18, 2026