Jupiter-G-8B
Jupiter-G-8B is a post-trained variant of Google Gemma 4 E4B IT, developed by Locai Labs. The G denotes the Gemma base. Jupiter-G-8B improves instruction following (+1.7 IFEval, +1.0 IFBench) and coding/agentic capability (+3.2 LCB pass@1) while preserving the base model's reasoning and knowledge through our Forget-Me-Not™ framework. This model was trained on 1 H200 GPU using 100% renewable energy.
Benchmarks
We evaluate Jupiter-G-8B against gemma-4-E4B-it.
| Benchmark | Jupiter-G-8B | gemma-4-E4B-it |
|---|---|---|
| IFEval (prompt strict) | 89.3 | 87.6 |
| IFBench (prompt strict) | 35.4 | 34.4 |
| AgentHarm harm rate | 12.0 | 22.3 |
| MMLU Redux | 82.0 | 83.4 |
| LiveCodeBench v6 | 55.2 | 52.0 |
IFEval and IFBench both reported with prompt strict accuracy. LiveCodeBench v6 reported with pass@1.
Training
Post-Training Data
Jupiter-G-8B is fine-tuned on a curated mixture of five datasets:
| Dataset | Domain | N |
|---|---|---|
| Self-cognition (non-reasoning) | Identity | ~full |
| UltraChat (reasoning + non-reasoning) | Reasoning / Replay | 12,500 |
| Nemotron terminal trajectories (reasoning) | Terminal / Agentic | 20,000 |
| Nemotron competitive programming (non-reasoning) | Coding | 20,000 |
Training Configuration
| Method | LoRA (rank 16, alpha 32) |
|---|---|
| Target Modules | All linear layers |
| Epochs | 2 |
| Optimiser | AdamW (fused) |
| Learning rate | 2e-4 (cosine decay, 5% warmup) |
| Weight decay | 0.001 |
| Max grad norm | 1.0 |
| Batch size | 64 (global: 8 local x 8 accumulation) |
| Sequence length | 2,048 |
| Precision | BF16 |
| Gradient checkpointing | Enabled |
| Loss | Assistant-only |
| Kernel | Liger |
| Attention | Eager |
Key Techniques
- Forget-Me-Not: Synthetic replay data generated by the unmodified base model on UltraChat prompts, preserving existing capabilities during domain-specific fine-tuning.
- Agentic/terminal training: Curated terminal trajectories from NVIDIA's Nemotron-Terminal-Corpus.
- Competitive programming: Exercism-derived programming problems from Nemotron corpora to strengthen code generation.
Citation
@misc{locailabs2026jupiterg,
title = {Jupiter-G-8B},
author = {George Drayson},
year = {2026},
url = {https://huggingface.co/locailabs/Jupiter-G-8B}
}
Acknowledgements
Jupiter-G-8B builds on Google Gemma 4. Terminal and programming data are sourced from NVIDIA's Nemotron corpora.