AlienKevin's picture
Upload Nemotron-Terminal-8B reproduction weights (Marin SFT, exp3490b, step-5720)
16e666b verified
|
Raw
History Blame Contribute Delete
2.82 kB
metadata
license: apache-2.0
base_model: Qwen/Qwen3-8B
tags:
  - terminal
  - sft
  - marin
  - nemotron
  - qwen3
datasets:
  - nvidia/Nemotron-Terminal-Corpus
model-index:
  - name: nemotron-terminal-8b-repro
    results:
      - task:
          type: text-generation
          name: Terminal-Bench 2.0
        dataset:
          name: Terminal-Bench 2.0
          type: custom
        metrics:
          - type: accuracy
            value: 15.9
            name: TB2 Accuracy (%)
      - task:
          type: text-generation
          name: Terminal-Bench Lite
        dataset:
          name: OpenThoughts-TBLite
          type: open-thoughts/OpenThoughts-TBLite
        metrics:
          - type: accuracy
            value: 29
            name: TBLite Accuracy (%)

Nemotron-Terminal-8B Reproduction (Marin)

A reproduction of NVIDIA's Nemotron-Terminal-8B by the Marin community, SFT'd on the full Nemotron-Terminal-Corpus (366K examples).

Results

Benchmark Nemotron-Terminal-8B (released) This model (Marin SFT) Diff
Terminal-Bench 2.0 13.0% ± 2.2 15.9% (14/88) +2.9pp
Terminal-Bench Lite 23.0% (23/100) 29.0% (29/100) +6.0pp

Both benchmarks show our reproduction outperforming the released weights, despite training on 366K examples vs the paper's 490K.

Training Details

  • Base model: Qwen/Qwen3-8B
  • Training data: nvidia/Nemotron-Terminal-Corpus — 366K examples (skill_based_easy, skill_based_medium, skill_based_mixed, diverse_complex subsets)
  • Steps: 5,721 (2 epochs)
  • Learning rate: 2e-5 (cosine schedule)
  • Batch size: 128
  • Sequence length: 32,768
  • Final loss: ~0.35
  • Hardware: TPU v5p-32
  • Framework: Marin (Levanter-based)
  • W&B run: exp3490b

Evaluation Setup

  • Inference: vLLM-TPU (native) on v5p-8
  • Agent: Terminus-2 (temperature=0.6)
  • Environments: Daytona

Links