Skip to main content

Command Line

eole is the single command line utility used to run various tools. These tools are categorized into several groups:

Categories of Tools​

Main Entrypoints​

  • build_vocab
  • train
  • predict
  • serve

Model Conversion Tools​

  • convert
    • Flavors: HF (supported Hugging Face architectures), T5, COMET, MetricX, onmt_config (legacy OpenNMT-py config migration)

Model Management Tools​

  • model
    • Subcommands: lora, release, extract_embeddings, average

Miscellaneous Tools (Mostly Legacy)​

  • tools
    • Subcommands: LM_scoring, oracle_comet, run_mmlu, spm_to_vocab, mbr_bleu, embeddings_to_torch, oracle_bleu, hellaswag

Usage​

The main entrypoints are typically used with a yaml configuration file. Top-level run parameters can be overridden with CLI flags; edit nested YAML fields directly. Dotted overrides are not supported.

Examples​

eole build_vocab -c your_config.yaml
eole train -c your_config.yaml
eole predict -c your_config.yaml

Additional Tools​

All other tools have specific arguments that can be inspected via the command helpstring -h.

Example​

eole convert HF -h
eole model lora -h
eole serve -h

Native EOLE COMET Scoring​

EOLE supports native scoring with pre-converted Unbabel COMET metrics through EOLE-COMET, EOLE-COMET-KIWI, and EOLE-XCOMET. These hosted EOLE models can be used directly for validation scoring or with eole predict --with_score. See the EOLE-COMET Hugging Face collection for all published COMET-family EOLE conversions.

eole predict \
  --model_path eole-nlp/wmt22-comet-da-eole-fp16 \
  --src /path/to/src.txt \
  --tgt /path/to/mt.txt \
  --ref /path/to/ref.txt \
  --output /path/to/scores.txt \
  --with_score

Use the hosted defaults in your training config with:

  • valid_metrics: ["EOLE-COMET"], valid_metrics: ["EOLE-COMET-KIWI"], or valid_metrics: ["EOLE-XCOMET"]

Set comet_model only when you want to override the default hosted EOLE model. Validation metrics configured with valid_metrics use comet_compute_dtype and default to fp16. Direct eole predict scoring uses the regular inference dtype knob; pass --compute_dtype fp32 if you need fp32 CLI scoring.

Raw Unbabel COMET repos still need conversion before EOLE can load them:

eole convert COMET --model Unbabel/wmt22-comet-da --output /path/to/converted/wmt22-comet-da
eole convert COMET --model Unbabel/wmt22-cometkiwi-da --output /path/to/converted/wmt22-cometkiwi-da
eole convert COMET --model Unbabel/XCOMET-XL --token "$HF_TOKEN" --output /path/to/converted/XCOMET-XL

See recipes/scoring/comet_native/ for hosted model IDs, direct scoring examples, conversion options, and parity checks.

Native EOLE MetricX Scoring​

EOLE supports native scoring with pre-converted Google MetricX models through EOLE-METRICX and EOLE-METRICX-QE. Both scorers default to the hosted fp32 EOLE model eole-nlp/metricx-24-hybrid-large-v2p6-eole. See the EOLE-MetricX Hugging Face collection for all published MetricX EOLE conversions.

eole predict \
  --model_path eole-nlp/metricx-24-hybrid-large-v2p6-eole \
  --src /path/to/src.txt \
  --tgt /path/to/mt.txt \
  --ref /path/to/ref.txt \
  --output /path/to/metricx-scores.txt \
  --with_score \
  --compute_dtype fp32

Omit --ref to score in QE mode. MetricX scores are lower-is-better raw error scores. Validation metrics configured with valid_metrics use metricx_compute_dtype and default to fp32 for stability. Direct eole predict scoring uses --compute_dtype.

Use the hosted defaults in your training config with:

  • valid_metrics: ["EOLE-METRICX"] for reference-based scoring
  • valid_metrics: ["EOLE-METRICX-QE"] for reference-free scoring

Set metricx_model only when you want to override the default hosted EOLE model. Raw Google MetricX repos still need conversion before EOLE can load them:

eole convert MetricX --model google/metricx-24-hybrid-large-v2p6 --dtype fp32 --output /path/to/converted/metricx-24-hybrid-large-v2p6-eole

See recipes/scoring/metricx_native/ for hosted model IDs, direct scoring examples, conversion options, and MetricX-23/24 input mode details.

Serving and speculative decoding​

Use the server recipe for YAML configuration and API examples. Qwen3.8/MTP and Claude Code document native drafting, verification, current parity limitations, and the local client setup.