File size: 8,140 Bytes
f713f1f
 
 
 
 
15ab6d3
f713f1f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
213
214
215
216
217
218
219
220
221
222
223
224
225
226
// Runs SmolLM2-360M-Instruct (q4f16, ~273 MB) on the WASM backend, off the main thread.
import * as tf from "https://cdn.jsdelivr.net/npm/@huggingface/transformers@3";

const { pipeline, TextStreamer, InterruptableStoppingCriteria } = tf;

const MODEL_ID = "HuggingFaceTB/SmolLM2-135M-Instruct";
let generator = null;

self.onmessage = async (e) => {
  const msg = e.data;
  try {
    if (msg.type === "load") {
      const progress_callback = (data) => {
        // Cache hits send non-cloneable payloads sometimes; keep it plain.
        const { status, file, progress, loaded, total } = data;
        self.postMessage({ type: "progress", data: { status, file, progress, loaded, total } });
      };
      // Prefer WebGPU when the adapter supports fp16 (q4f16 needs it);
      // fall back to WASM otherwise or if WebGPU init fails.
      let device = "wasm";
      try {
        const adapter = await self.navigator?.gpu?.requestAdapter?.();
        if (adapter?.features?.has("shader-f16")) device = "webgpu";
      } catch {}
      try {
        generator = await pipeline("text-generation", MODEL_ID, {
          dtype: "q4f16",
          device,
          progress_callback,
        });
      } catch (err) {
        if (device === "wasm") throw err;
        device = "wasm";
        generator = await pipeline("text-generation", MODEL_ID, {
          dtype: "q4f16",
          device,
          progress_callback,
        });
      }
      self.postMessage({ type: "ready", device });
    } else if (msg.type === "condense") {
      const line = await condensePair(msg.pair, msg.before, msg.after);
      self.postMessage({ type: "result", id: msg.id, line });
    } else if (msg.type === "expand") {
      const lines = await expandLine(msg.line, msg.before, msg.after);
      self.postMessage({ type: "result", id: msg.id, lines });
    }
  } catch (err) {
    self.postMessage({ type: "error", id: msg.id, message: err.message ?? String(err) });
  }
};

// Apply the chat template ourselves, then append `prefill` so the model's first
// generated token is already inside the JSON value β€” it cannot add a preamble.
// `stopPattern` interrupts generation as soon as the JSON value is closed.
async function generateConstrained(messages, prefill, stopPattern, maxNewTokens) {
  const tokenizer = generator.tokenizer;
  const prompt =
    tokenizer.apply_chat_template(messages, { tokenize: false, add_generation_prompt: true }) +
    prefill;

  let acc = "";
  const opts = {
    max_new_tokens: maxNewTokens,
    do_sample: true,
    temperature: 0.7,
    top_p: 0.9,
    repetition_penalty: 1.15,
    no_repeat_ngram_size: 3,
    return_full_text: false,
  };
  if (typeof InterruptableStoppingCriteria === "function" && typeof TextStreamer === "function") {
    const stopper = new InterruptableStoppingCriteria();
    opts.stopping_criteria = stopper;
    opts.streamer = new TextStreamer(tokenizer, {
      skip_prompt: true,
      skip_special_tokens: true,
      callback_function: (text) => {
        acc += text;
        if (stopPattern.test(acc)) stopper.interrupt();
      },
    });
  }
  const out = await generator(prompt, opts);
  return out[0].generated_text;
}

function unescapeJson(s) {
  try {
    return JSON.parse('"' + s + '"');
  } catch {
    return s;
  }
}

function tidy(s) {
  return s
    .replace(/\s+/g, " ")
    // JSON debris the parse fallback can drag along, e.g. a trailing '} or "]}.
    .replace(/["'’\s]*[\]}]+["'’\s]*$/, "")
    .trim();
}

// If the model looped and never closed the quote, the parse fallback swallows
// the whole ramble. Cut an overlong line back at the seam nearest the budget.
function clampLine(line, budget) {
  if (line.length <= budget) return line;
  const seams = [...line.matchAll(/[,;.—–]\s+|\s+(?:and|but|or|while|when|so)\s+/g)]
    .filter((m) => m.index > 0 && m.index <= budget);
  if (seams.length) return line.slice(0, seams.at(-1).index).trim();
  const cut = line.lastIndexOf(" ", budget);
  return line.slice(0, cut > 0 ? cut : budget).trim();
}

function contextNote(before, after) {
  let note = "";
  if (before?.length) note += `\nFor context, the lines just before them:\n${before.join("\n")}\n`;
  if (after?.length) note += `\nFor context, the lines just after them:\n${after.join("\n")}\n`;
  return note;
}

// A literal placeholder like {"line": "your new line here"} gets parroted
// verbatim by a 360M model, so the format is taught with a real few-shot
// exchange instead β€” the condense and expand examples are inverses.
const EXAMPLE_TWO = ["the rain fell all night on the tin roof", "and we listened without speaking"];
const EXAMPLE_ONE = "all night we listened to the rain speak on the tin roof";

async function condensePair([a, b], before, after) {
  const messages = [
    {
      role: "system",
      content:
        "You are a poet who condenses poems. You reply with exactly one JSON object and nothing else.",
    },
    {
      role: "user",
      content: `Two consecutive lines of a poem:

Line A: ${EXAMPLE_TWO[0]}
Line B: ${EXAMPLE_TWO[1]}

Combine Line A and Line B into ONE short poetic line that keeps their imagery, meaning, and voice. Reply with JSON only.`,
    },
    {
      role: "assistant",
      content: `{"line": "${EXAMPLE_ONE}"}`,
    },
    {
      role: "user",
      content: `Two consecutive lines of a poem:

Line A: ${a}
Line B: ${b}
${contextNote(before, after)}
Combine Line A and Line B into ONE short poetic line that keeps their imagery, meaning, and voice. Reply with JSON only.`,
    },
  ];

  // Generation continues from `{"line": "` β€” capture up to the first unescaped quote.
  const raw = await generateConstrained(messages, '{"line": "', /(?:^|[^\\])"/, 60);
  const m = raw.match(/^((?:[^"\\]|\\.)*)"/s);
  const line = clampLine(
    tidy(unescapeJson(m ? m[1] : raw.split("\n")[0])),
    Math.max(50, a.length + b.length),
  );
  // A condensation that vanished is worse than no condensation: keep the originals.
  return line || tidy(`${a} ${b}`);
}

// Split one string into two lines at a natural seam (for fallback parsing).
function splitInTwo(s) {
  const mid = s.length / 2;
  const seams = [...s.matchAll(/[,;—–]\s+|\s+(?:and|but|or|while|when|so)\s+/g)];
  if (seams.length) {
    const best = seams.reduce((p, c) => (Math.abs(c.index - mid) < Math.abs(p.index - mid) ? c : p));
    return [s.slice(0, best.index + (best[0].startsWith(" ") ? 0 : 1)).trim(), s.slice(best.index + best[0].length).trim()];
  }
  const words = s.split(" ");
  const cut = Math.ceil(words.length / 2);
  return [words.slice(0, cut).join(" "), words.slice(cut).join(" ")];
}

async function expandLine(line, before, after) {
  const messages = [
    {
      role: "system",
      content:
        "You are a poet who elaborates poems. You reply with exactly one JSON object and nothing else.",
    },
    {
      role: "user",
      content: `One line of a poem:

Line: ${EXAMPLE_ONE}

Rewrite this single line as TWO short poetic lines that together carry the same meaning, imagery, and voice. Reply with JSON only.`,
    },
    {
      role: "assistant",
      content: `{"lines": ["${EXAMPLE_TWO[0]}", "${EXAMPLE_TWO[1]}"]}`,
    },
    {
      role: "user",
      content: `One line of a poem:

Line: ${line}
${contextNote(before, after)}
Rewrite this single line as TWO short poetic lines that together carry the same meaning, imagery, and voice. Reply with JSON only.`,
    },
  ];

  // Generation continues from `{"lines": ["` β€” stop once the array closes.
  const raw = await generateConstrained(messages, '{"lines": ["', /"\s*\]/, 90);
  const two = raw.match(/^((?:[^"\\]|\\.)*)"\s*,\s*"((?:[^"\\]|\\.)*)/s);
  const budget = Math.max(60, line.length * 2);
  if (two) {
    const first = clampLine(tidy(unescapeJson(two[1])), budget);
    const second = clampLine(tidy(unescapeJson(two[2])), budget);
    if (first && second) return [first, second];
  }
  const one = raw.match(/^((?:[^"\\]|\\.)*)"/s);
  const text = clampLine(tidy(unescapeJson(one ? one[1] : raw.split("\n")[0])), budget * 2);
  if (!text) return [tidy(line)];
  return splitInTwo(text);
}