Command-line Interface

The inif CLI for converting eval logs and viewing INIF files in the browser.

The inif CLI exposes the same converters and viewer that the Python API does. It is the right tool for one-off conversions, batch jobs, and quick-look browsing.

inif --help

Subcommands

Command	Purpose
`inif view <input>`	Open an INIF document in the HTML viewer.
`inif convert txt <inputs>`	Tokenize raw text files / strings.
`inif convert eval <inputs>`	Convert Inspect AI `.eval` logs.
`inif convert evaleval <input>`	Convert every_eval_ever instance records.

All converters write .inif.json by default. To produce the indexed archive instead, give an output path with the .inif suffix (-o traces.inif).

`inif view`

inif view traces.inif.json                 # opens in default browser
inif view traces.inif -o traces.html       # writes self-contained HTML
inif view traces.inif.json --compact       # hide empty panels
inif view traces.inif.json --title "Run 42"

When -o is omitted, the rendered HTML is written to a temp file and opened in your default browser. With -o, the file is written to that path and not opened.

`inif convert txt`

Tokenize raw text files (or directories) with a HuggingFace tokenizer.

inif convert txt input.txt -m gpt2
inif convert txt corpus/ -m gpt2 --glob "*.md"
inif convert txt a.txt b.txt c.txt -m gpt2 -o batch.inif.json

Each input file becomes one Sample; sample ids default to the filename. Sequence deduplication runs by default — pass --no-dedup to skip it.

Flag	Default	Effect
`-m, --model`	required	Tokenizer model id (passed to `AutoTokenizer.from_pretrained`).
`--revision`	unset	Pinned model revision. Stamped into `metadata.model.revision`.
`--min-seq-length`	`5`	Minimum n-gram length for sequence dedup.
`--glob`	`*.txt`	When an input is a directory, glob pattern to filter files.
`--no-dedup`	off	Disable sequence deduplication.
`-o, --output`	first input’s name with `.inif.json`	Output path.

`inif convert eval`

Convert one or more Inspect AI .eval logs.

inif convert eval logs/run.eval
inif convert eval logs/*.eval -o batch.inif.json
inif convert eval logs/run.eval -m meta-llama/Llama-3.1-8B
inif convert eval logs/run.eval --no-tag-chat-roles --no-dedup

Flag	Default	Effect
`-m, --model`	autodetect	Tokenizer override. By default, the tokenizer is auto-loaded from the eval’s model id (with routing prefixes like `together/` stripped). See Tokenizer resolution.
`--revision`	unset	Pin `metadata.model.revision`.
`--min-seq-length`	`5`	Sequence dedup minimum.
`--no-tag-chat-roles`	off	Skip chat-template role annotations.
`--no-dedup`	off	Skip sequence deduplication.
`-o, --output`	input stem with `.inif.json`	Output path.

When multiple inputs are given, each writes to its own output file (the -o flag applies to all of them — useful when you only have one input).

`inif convert evaleval`

Convert every_eval_ever instance-level records.

# Local JSON / JSONL
inif convert evaleval instances.jsonl
inif convert evaleval instances.json --aggregate eval.json

# From the EEE_datastore on HuggingFace
inif convert evaleval --hf-config theory_of_mind_samples
inif convert evaleval --hf-config theory_of_mind_samples \
                     --hf-aggregate-config theory_of_mind \
                     --limit 100

Flag	Default	Effect
`input`	(positional)	Path to instance records (`.json` or `.jsonl`). Omit when using `--hf-config`.
`--aggregate`	unset	Optional aggregate `eval.json` for document-level metadata.
`--hf-config`	unset	Config name from `evaleval/EEE_datastore`.
`--hf-split`	`samples`	HF split for the records.
`--hf-repo`	`evaleval/EEE_datastore`	HF dataset repo (override for testing).
`--hf-aggregate-config`	unset	Paired aggregate config (non-`_samples` counterpart).
`--limit`	unset	Cap the number of records converted.
`-m, --model`	autodetect	Tokenizer override. Same rules as `convert eval`.
`--revision`	unset	Pin `metadata.model.revision`.
`--min-seq-length`	`5`	Sequence dedup minimum.
`--no-tag-chat-roles`	off	Skip chat-template role annotations.
`--no-dedup`	off	Skip sequence deduplication.
`-o, --output`	derived from input or HF config	Output path.

The HF route streams records (it does not download the entire dataset locally) and respects --limit while iterating.

Exit codes

The CLI exits with status 1 and prints a message to stderr when no input files match (inif convert txt). All other failures bubble up as raw Python tracebacks — the CLI is intentionally a thin wrapper around the Python API.