File size: 8,092 Bytes
be1ce5d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
#!/usr/bin/env python3
"""
build_dataset.py — Turn cleaned entries into a dual-mode training dataset.

Modes taught simultaneously (control-token dropout):
  * PERSONA mode: a persona token (<|pepys|> etc.) precedes the entry; the model
    learns conditional voice generation.
  * BLENDED mode: no persona token; the model marginalizes over all four ghosts
    and a composite voice emerges.
Each entry is emitted with the persona token present with probability
--persona-token-prob (default 0.5).

Entry document format (one training document per entry):

    <|entry|><|maclane|>
    19 January 1917.
    <text...>
    <|/entry|>

A fraction of examples (--prompt-frac) carry a generic introspection prompt:

    <|entry|><|mansfield|>
    [Prompt: What do you keep refusing to look at?]
    4 March 1920.
    <text...>
    <|/entry|>

Inference recipes (see sample.py):
    blended, fresh:   "<|entry|>\n7 June 2026.\n"
    persona, seeded:  "<|entry|><|vangogh|>\n7 June 2026.\nToday was hard."
    prompted:         "<|entry|>\n[Prompt: <your prompt>]\n7 June 2026.\n"

Balancing: Pepys is ~10x the other corpora; entries are stratified-sampled by
year down to --pepys-char-budget characters so one ghost doesn't possess the
other three.

Usage: python scripts/build_dataset.py [--seed 42] [--val-frac 0.02] ...
"""
import argparse
import json
import random
import re
from collections import defaultdict
from pathlib import Path

PERSONAS = ["pepys", "vangogh", "mansfield", "maclane"]
ENTRY_OPEN, ENTRY_CLOSE = "<|entry|>", "<|/entry|>"
PERSONA_TOKENS = {p: f"<|{p}|>" for p in PERSONAS}
SPECIAL_TOKENS = [ENTRY_OPEN, ENTRY_CLOSE, *PERSONA_TOKENS.values()]

# Generic introspection prompts: deliberately answerable by *any* diary entry,
# formatted as first-person rhetorical questions written by the diarist themselves.
GENERIC_PROMPTS = [
    "What is on my mind today?",
    "I must write about my day.",
    "What exactly happened today?",
    "How am I feeling, honestly?",
    "What do I keep returning to?",
    "I need to set down the truth of this day.",
    "What did I notice today that no one else did?",
    "What am I afraid of right now?",
    "What do I actually want?",
    "I will write until something true appears.",
    "What would I say if no one could ever read this?",
    "Describe where I am.",
    "What is the weather inside me?",
    "What am I working on, and how does it go?",
    "Who occupied my thoughts today?",
    "What small thing mattered today?",
    "I must confess something.",
    "What do I keep refusing to look at?",
    "I must make an account of myself.",
    "What does this day deserve to have remembered of it?",
]


def normalize_date(date: str | None) -> str:
    if not date:
        return "An unmarked day"
    d = re.sub(r"(\d+)(st|nd|rd|th)", r"\1", date)  # 16th -> 16
    return d.strip()


def format_doc(entry: dict, use_persona: bool, prompt: str | None) -> str:
    parts = [ENTRY_OPEN]
    if use_persona:
        parts.append(PERSONA_TOKENS[entry["persona"]])
    parts.append("\n")
    parts.append(f"{normalize_date(entry['date'])}.\n")
    if prompt:
        parts.append(f"{prompt}\n\n")
    if entry.get("title"):
        parts.append(f"{entry['title']}\n")
    parts.append(entry["text"].strip())
    parts.append(f"\n{ENTRY_CLOSE}")
    return "".join(parts)


def subsample_pepys(entries: list[dict], char_budget: int, rng: random.Random) -> list[dict]:
    """Stratified by year so the whole 1660s decade survives the cut."""
    by_year = defaultdict(list)
    for e in entries:
        m = re.search(r"(\d{4})", e["date"] or "")
        by_year[m.group(1) if m else "?"].append(e)
    for v in by_year.values():
        rng.shuffle(v)
    picked, used = [], 0
    # round-robin across years until budget exhausted
    pools = list(by_year.values())
    i = 0
    while used < char_budget and any(pools):
        pool = pools[i % len(pools)]
        if pool:
            e = pool.pop()
            picked.append(e)
            used += len(e["text"])
        i += 1
        if all(not p for p in pools):
            break
    return picked


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--clean-dir", default="data/clean", type=Path)
    ap.add_argument("--out-dir", default="data/dataset", type=Path)
    ap.add_argument("--seed", type=int, default=42)
    ap.add_argument("--persona-token-prob", type=float, default=0.5)
    ap.add_argument("--prompt-frac", type=float, default=0.15)
    ap.add_argument("--val-frac", type=float, default=0.02)
    ap.add_argument("--pepys-char-budget", type=int, default=900_000,
                    help="Max characters of Pepys to keep (he is 10x the others raw)")
    ap.add_argument("--chunk-prob", type=float, default=0.25,
                    help="Probability of grouping sequential entries into a multi-day chunk")
    ap.add_argument("--chunk-max-size", type=int, default=4,
                    help="Max number of entries in a temporal chunk")
    args = ap.parse_args()
    rng = random.Random(args.seed)
    args.out_dir.mkdir(parents=True, exist_ok=True)

    docs_train, docs_val, stats = [], [], {}
    for p in PERSONAS:
        entries = [json.loads(l) for l in (args.clean_dir / f"{p}.jsonl").open()]
        
        # Tag original order to restore chronological order after subsampling
        for idx, e in enumerate(entries):
            e["_idx"] = idx
            
        if p == "pepys":
            entries = subsample_pepys(entries, args.pepys_char_budget, rng)
            entries.sort(key=lambda x: x["_idx"])

        chunks = []
        i = 0
        while i < len(entries):
            if rng.random() < args.chunk_prob:
                chunk_sz = rng.randint(2, args.chunk_max_size)
            else:
                chunk_sz = 1
                
            chunk = entries[i:i+chunk_sz]
            if not chunk:
                break
                
            use_persona = rng.random() < args.persona_token_prob
            prompt = rng.choice(GENERIC_PROMPTS) if rng.random() < args.prompt_frac else None
            
            formatted_texts = []
            for j, e in enumerate(chunk):
                formatted_texts.append(format_doc(e, use_persona, prompt if j == 0 else None))
                
            chunks.append({
                "text": "".join(formatted_texts),
                "persona": p,
                "has_persona_token": use_persona,
                "has_prompt": prompt is not None,
                "num_entries": len(chunk)
            })
            i += chunk_sz
            
        rng.shuffle(chunks)
        n_val = max(2, int(len(chunks) * args.val_frac))
        val, train = chunks[:n_val], chunks[n_val:]
        docs_train.extend(train)
        docs_val.extend(val)
        
        chars = sum(len(e["text"]) for e in entries)
        stats[p] = {"entries": len(entries), "chunks": len(chunks), "chars": chars,
                    "train_chunks": len(train), "val_chunks": len(val)}
        print(f"[dataset] {p:10s} kept={len(entries):4d} entries -> {len(chunks)} chunks "
              f"chars={chars:8,d} (train {len(train)} / val {len(val)})")

    rng.shuffle(docs_train)
    for name, docs in (("train", docs_train), ("val", docs_val)):
        with (args.out_dir / f"{name}.jsonl").open("w") as f:
            for d in docs:
                f.write(json.dumps(d, ensure_ascii=False) + "\n")

    n_tok = sum(s["chars"] for s in stats.values()) // 4
    meta = {"personas": PERSONAS, "special_tokens": SPECIAL_TOKENS,
            "persona_token_prob": args.persona_token_prob,
            "prompt_frac": args.prompt_frac, "seed": args.seed,
            "chunk_prob": args.chunk_prob, "chunk_max_size": args.chunk_max_size,
            "approx_tokens": n_tok, "stats": stats,
            "generic_prompts": GENERIC_PROMPTS}
    (args.out_dir / "meta.json").write_text(json.dumps(meta, indent=2))
    print(f"[dataset] train={len(docs_train)} val={len(docs_val)} docs, "
          f"~{n_tok:,} tokens -> {args.out_dir}")

if __name__ == "__main__":
    main()