Grug 67B Nemotron Terminal SFT reproduction
This is the BF16 Hugging Face/vLLM export evaluated by the Marin experiment
sft-repro-nemotron-terminal-on-thinking-ckpt.
Training
- Initialization: Marin Grug 67B stage-2 thinking checkpoint, step 630
- Data:
nvidia/Nemotron-Terminal-Corpusat revisiona1667c4, using Marin'sdocuments/nemotron_terminal_full@2026.06.28selection andtokenized/nemotron_terminal_full-chat-7adc64@2026.07.17 - Steps: 1,888
- Sequence length: 32,768
- Global batch size: 64
- Peak learning rate: 5e-6, AdamH/Adam with cosine decay, 3% warmup, and 0.1 minimum learning-rate ratio
- Training geometry: 64 H100 GPUs; replica=1, data=8, expert=8, model=1
- Marin source:
e502edd70c
The training run completed with a final reported loss of 0.501. The export was generated from the final step-1,888 checkpoint and preserves the model's native 32,768-token context limit.
Evaluation
The model was evaluated on the 300-trial OpenThoughts-TBLite reproduction campaign with the Terminus-2 harness. The experimental controls use three attempts per task, temperature 0.6, and an agent context budget of 24,576 input plus 8,192 output tokens. Evaluation results are tracked separately from this weight artifact.
Provenance
The immutable source export is:
s3://marin-us-east-02a/marin/grug/sft_repro_thinking_step630_nemotron_terminal_steps1888/2026.08.11.1/exports/step-1888/hf-bf16-vllm/