ikhou/dict-xs-mlc-q4f16_1

Ikhou Dict XS (MLC/WebLLM)

Ikhou Dict XS (MLC) is a WebGPU-ready, quantized build of ikhou/dict-xs for in-browser inference with MLC-LLM / WebLLM. The source model is a Qwen/Qwen3-0.6B fine-tune for bilingual dictionary glossing.

Model Summary

• Task: single-line dictionary gloss generation
• Format: chat-based (system + user + assistant)
• Quantization: q4f16_1 (4-bit weights, fp16 activations)
• Target runtime: WebLLM (WebGPU)

Training Data (Source Model)

The source model ikhou/dict-xs was trained on synthetic bilingual dictionary examples generated from multilingual web corpora (FineWeb-2 with optional FineWeb). The pipeline:

1. Extract sentences from multilingual corpora.
2. Select a word or short phrase from each sentence.
3. Sample a target language (cross-lingual by default).
4. Use a teacher LLM to generate a short gloss under a strict rubric.
5. Filter, de-duplicate, and write SFT JSONL with a system prompt, user prompt, and <final>...</final> assistant response.

See ikhou/dict-xs for full training details.

Prompt Format

System prompt used during training:

You are a bilingual dictionary assistant.

Your job: given a word/phrase in context, output a SHORT dictionary-style gloss line.

Hard rules:
- Output EXACTLY ONE LINE and nothing else.
- No quotes, no bullets, no labels (no "Definition:", "Meaning:", etc).
- Do NOT repeat the original word/phrase in the output.
- Keep it short (ideally <= 120 characters).

Gloss rules:
- Output 1-4 translations/synonyms in the definition language, separated by ", ".
- Each gloss should be short (1-3 words). Prefer common, user-friendly glosses.
- Do NOT write full sentences. No trailing period.

French grammar hints (only if confident):
IMPORTANT: The French-only formatting hints below apply ONLY when the definition language is French (fr/fra).
If the definition language is NOT French, do NOT use nm./nf./adj./adv., do NOT add French tense notes, and do NOT add (pp).
- Noun: prefix with "nm." (masc) or "nf." (fem), then a space, then glosses.
  Example: nm. face
- Adjective: prefix with "adj.", then a space, then glosses.
  Example: adj. fragile, delicate
- Adverb: prefix with "adv.", then a space, then glosses.
  Example: adv. extremely, exceedingly
- Conjugated verb form: glosses, then add "(tense, subject)" in French.
  Example: came back, used to come back (imparfait, il)
- Past participle: glosses, then add "(pp)".
  Example: watched over, supervised (pp)

User prompt template:

Expression: "<selection>"
Context: <sentence>
Source language: <iso639-3> (<name>)
Definition language: <iso639-3> (<name>)

Return the single-line gloss now.

WebLLM Usage

import { MLCEngine } from "@mlc-ai/web-llm";

const engine = new MLCEngine();
await engine.reload("ikhou/dict-xs-mlc-q4f16_1");

const messages = [
  { role: "system", content: "<system prompt above>" },
  {
    role: "user",
    content:
      "Expression: \"online\"\n" +
      "Context: He paid for the course online and started immediately.\n" +
      "Source language: eng (English)\n" +
      "Definition language: spa (Spanish)\n\n" +
      "Return the single-line gloss now.",
  },
];

const reply = await engine.chat.completions.create({
  messages,
  temperature: 0.0,
  max_tokens: 64,
});

console.log(reply.choices[0].message.content);

MLC Build Notes

This repository is produced from ikhou/dict-xs using MLC-LLM:

mlc_llm convert_weight /path/to/ikhou/dict-xs \
  --quantization q4f16_1 \
  -o ./dict-xs-mlc-q4f16_1

mlc_llm gen_config /path/to/ikhou/dict-xs \
  --quantization q4f16_1 \
  --conv-template qwen2 \
  -o ./dict-xs-mlc-q4f16_1

mlc_llm compile ./dict-xs-mlc-q4f16_1/mlc-chat-config.json \
  --device webgpu \
  -o ./dict-xs-mlc-q4f16_1/dict-xs-q4f16_1-webgpu.wasm

If your MLC version uses a different conversation template name for Qwen3, replace qwen2 with the appropriate template.

Limitations

• Outputs can be inaccurate or inconsistent with the rubric.
• The model inherits biases from source corpora and the teacher model.
• Specialized domains and low-resource languages may be weaker.

Acknowledgements

Base model: Qwen/Qwen3-0.6B. MLC runtime: MLC-LLM / WebLLM.