Skip to main content

apr_cli/
tokenize_commands.rs

1/// Tokenizer training pipeline subcommands (forjar-style plan/apply).
2///
3/// Thin CLI wrappers around aprender's BPE training infrastructure.
4/// Trains a BPE vocabulary from a text corpus for use in model training.
5#[derive(Subcommand, Debug)]
6pub enum TokenizeCommands {
7    /// Validate inputs and estimate tokenizer training time/resources.
8    ///
9    /// Checks that the input corpus exists, counts lines/bytes, estimates
10    /// vocabulary coverage, and reports expected training time. Outputs a
11    /// serializable plan manifest (text, JSON, or YAML).
12    ///
13    /// Analogous to `forjar plan` — shows what will happen before committing.
14    Plan {
15        /// Path to training corpus (text file, one document per line)
16        #[arg(long, value_name = "FILE")]
17        data: PathBuf,
18        /// Target vocabulary size
19        #[arg(long, default_value = "32000")]
20        vocab_size: usize,
21        /// Tokenizer algorithm: bpe, wordpiece, unigram
22        #[arg(long, default_value = "bpe")]
23        algorithm: String,
24        /// Output directory for trained tokenizer
25        #[arg(short, long, default_value = "./tokenizer-output")]
26        output: PathBuf,
27        /// Output format: text, json, yaml
28        #[arg(long, default_value = "text")]
29        format: String,
30    },
31
32    /// Train a tokenizer on the corpus.
33    ///
34    /// Reads the input corpus, trains a BPE/WordPiece/Unigram tokenizer,
35    /// and writes vocab.json + merges.txt to the output directory.
36    ///
37    /// Analogous to `forjar apply` — commits resources and executes the plan.
38    Apply {
39        /// Path to training corpus (text file, one document per line)
40        #[arg(long, value_name = "FILE")]
41        data: PathBuf,
42        /// Target vocabulary size
43        #[arg(long, default_value = "32000")]
44        vocab_size: usize,
45        /// Tokenizer algorithm: bpe, wordpiece, unigram
46        #[arg(long, default_value = "bpe")]
47        algorithm: String,
48        /// Output directory for trained tokenizer
49        #[arg(short, long, default_value = "./tokenizer-output")]
50        output: PathBuf,
51        /// Maximum number of lines to read from corpus (0 = all)
52        #[arg(long, default_value = "0")]
53        max_lines: usize,
54    },
55
56    /// Train BPE on a JSONL corpus per contracts/tokenizer-bpe-v1.yaml (MODEL-2).
57    ///
58    /// Walks `--corpus` (file or directory of `.jsonl` files), extracting the
59    /// `content` field from each line, applies `--normalization` (NFC default),
60    /// and trains a BPE tokenizer with the target vocab size. Writes
61    /// `vocab.json` (token→id) and `merges.txt` (one `a b` pair per line, in
62    /// merge order) to `--output`.
63    Train {
64        /// Path to corpus: a `.jsonl` file or a directory containing `.jsonl` files.
65        /// Each line must be a JSON object with a `content` field.
66        #[arg(long, value_name = "PATH")]
67        corpus: PathBuf,
68        /// Target vocabulary size. Default 50_257 matches GPT-2 convention
69        /// (50_000 BPE merges + 256 byte-level fallback tokens + 1 sentinel)
70        /// and the MODEL-2 albor tokenizer contract (tokenizer-bpe-v1 v1.2.0).
71        #[arg(long, default_value = "50257")]
72        vocab_size: usize,
73        /// Minimum frequency a byte-pair must reach before BPE merges it into
74        /// a new vocabulary token (honored by `entrenar::tokenizer::BPETokenizer`
75        /// per task #103). Pairs below this threshold are left unmerged —
76        /// contract INV-TOK-002 of `contracts/tokenizer-bpe-v1.yaml`.
77        #[arg(long, default_value = "2")]
78        min_frequency: usize,
79        /// Output directory; will contain vocab.json and merges.txt.
80        #[arg(long, default_value = "./tokenizer-output")]
81        output: PathBuf,
82        /// Unicode normalization form applied to each document before training.
83        #[arg(long, default_value = "nfc")]
84        normalization: String,
85    },
86
87    /// Import a HuggingFace tokenizer.json into aprender's two-file
88    /// vocab.json + merges.txt layout per
89    /// `contracts/apr-cli-tokenize-import-hf-v1.yaml` (§50.4 step 5g.0).
90    ///
91    /// Reads `<INPUT>` (a HF tokenizer.json with `model.type == "BPE"`),
92    /// extracts `model.vocab` → `<OUTPUT>/vocab.json`, `model.merges` →
93    /// `<OUTPUT>/merges.txt` (one space-separated merge per line), and
94    /// writes `<OUTPUT>/manifest.json` with extraction provenance
95    /// (source path, sha256, vocab_size, merges_count, timestamp).
96    ///
97    /// Non-BPE inputs (Unigram, WordPiece) are rejected fail-fast with a
98    /// clear error citing the contract id.
99    ///
100    /// Unblocks fine-tuning from public HF checkpoints (Qwen2.5/Llama2/
101    /// Mistral) which distribute as a single tokenizer.json. The output
102    /// dir is consumable by `apr tokenize encode-corpus --tokenizer <DIR>`
103    /// and `apr pretrain --tokenizer <DIR>` without modification.
104    ImportHf {
105        /// Path to input HuggingFace tokenizer.json (BPE model required).
106        #[arg(long, value_name = "FILE")]
107        input: PathBuf,
108        /// Output directory; will contain vocab.json + merges.txt + manifest.json.
109        #[arg(long, value_name = "DIR")]
110        output: PathBuf,
111        /// Include `added_tokens` in vocab.json (default: BPE state machine only).
112        /// Use this when the downstream consumer needs special tokens (e.g.,
113        /// `<|im_start|>`, `<|endoftext|>`) materialized in vocab.json itself.
114        #[arg(long, default_value_t = false)]
115        include_added_tokens: bool,
116    },
117
118    /// Encode a JSONL corpus into `.bin` shards per contracts/pretokenize-bin-v1.yaml.
119    ///
120    /// Loads a trained BPE tokenizer (vocab.json + merges.txt) from `--tokenizer`,
121    /// reads `--corpus` (file or directory of `.jsonl` files), encodes the
122    /// `--content-field` of each line to u32 tokens, and writes
123    /// `shard-NNNN.bin` files (flat little-endian u32 streams) into `--output`.
124    /// The output format is precisely what `ShardBatchIter` (aprender-train)
125    /// expects at MODEL-2 pretrain read time.
126    ///
127    /// Root-cause fix for the pretokenize-to-bin gap documented in
128    /// memory/project_shard_reader_bin_format.md — replaces a Python shim
129    /// that was flagged as MUDA on 2026-04-19.
130    #[cfg(feature = "training")]
131    EncodeCorpus {
132        /// Path to JSONL corpus file or directory of `.jsonl` files.
133        #[arg(long, value_name = "PATH")]
134        corpus: PathBuf,
135        /// Directory containing vocab.json + merges.txt from `apr tokenize train`.
136        #[arg(long, value_name = "DIR")]
137        tokenizer: PathBuf,
138        /// Output directory for shard-NNNN.bin + manifest.json.
139        #[arg(long, value_name = "DIR")]
140        output: PathBuf,
141        /// Target tokens per shard (shard closes once this limit is reached).
142        #[arg(long, default_value = "10000000")]
143        shard_tokens: usize,
144        /// JSONL field to encode (default: `content`).
145        #[arg(long, default_value = "content")]
146        content_field: String,
147        /// Unicode normalization (must match tokenizer training).
148        #[arg(long, default_value = "nfc")]
149        normalization: String,
150        /// EOS insertion policy: none|between|after.
151        #[arg(long, default_value = "between")]
152        eos_policy: String,
153    },
154}