Skip to content

Output and reports

The output directory holds a ready-to-load tokenizer (and model, if trimmed), plus two files that record the run:

  • trim_report.json — sizes, reduction, corpus coverage, per-reason token counts, verification results, parameters removed. This is a serialised TrimReport.
  • _trimbed_config.yaml — the fully resolved configuration, for provenance.

TrimPipeline.run returns the same TrimReport it writes, so a caller never has to read the file back.

The rendered summary

TrimReport.render prints a summary in this shape. The figures below are illustrative; your corpus and settings decide the actual numbers.

model            codefuse-ai/F2LLM-v2-160M
tokenizer type   BPE
vocabulary       151,669 -> 32,412 (78.6% removed, 282 structural)
corpus           200,000 docs, 91,204,331 tokens, 99.9612% covered by the kept vocabulary
architecture     Qwen3Model
embeddings       151,936 -> 32,448 rows, 122,355,712 of 160,384,000 parameters removed (76.3%)
verification     512/512 identical, 512/512 decode-equivalent, 1.0004x tokens
model check      passed on 8 texts of 40,960 tokens: max |dh| 2.38e-07, max |dlogit| n/a (tolerance 1e-05)
kept by          chat_template=41, corpus=32,004, dependency=126, structural=282
output           trimmed/f2llm-nl

from trimbed import TrimConfig, TrimPipeline

report = TrimPipeline(TrimConfig.from_yaml("my_config.yaml")).run()
print(report.render())
print(report.vocabulary.trimmed_size)

Reading the sections

vocabulary is a VocabularyReport. Its kept_by_reason is the provenance breakdown from Selection: an id kept for two reasons is counted under both, so the numbers do not sum to the vocabulary size. Two fields are worth checking on every run:

  • unknown_requested_tokens — token strings you asked for that this vocabulary does not contain. Usually a keep-list written for a different checkpoint.
  • dropped_requested_tokens — tokens that were requested but that max_vocab_size had to remove anyway, mapped to the reasons they had been requested.

corpus is a CorpusReport, where distinct_tokens is the ceiling on what a corpus-only selection can keep, and coverage is the share of corpus occurrences the kept vocabulary still covers.

model is a ModelReport, present only when trim_model was on. verification and model_verification are the two report models described in Trimming a model.

Dry runs

--dry-run (or dry_run on the pipeline) runs the whole selection and produces the same report without writing anything. That is what makes a sweep over selection.top_k values cheap. See Configuration.