Thai Meeting-Minutes Summarization

🥈 2nd place — AI Benchmark 2026, a national Thai-language NLP competition organized by NECTEC / NSTDA.

A retrieval-augmented summarization system for Thai parliamentary meeting documents. Given a long meeting document and a query, it produces an abstractive Thai answer together with the supporting paragraph references — output as submission.csv (ID, abstractive, refs).

Scoring

Submissions are ranked by a composite metric:

score = 0.45 × SS-score (semantic similarity, bge-m3)
      + 0.35 × RougeL   (answer overlap)
      + 0.20 × IoU       (reference set match)

Approach

The winning configuration is a two-stage LLM pipeline over the whole document (an early finding — exp35/exp37 — was that feeding the full doc to a strong LLM beats every retrieval setup, so the final system has no retrieval step):

• Stage A — reference picker: an NVFP4-quantized Gemma 26B model reads the full document and selects the supporting paragraphs (reference IoU ≈ 0.82). Its chosen indices also become a hint passed to Stage B.
• Stage B — answer writer: an AWQ-quantized Qwen3-32B model writes the final Thai abstractive answer, citing paragraphs inline as [อ้างอิง: X] and using Stage A's hint. Context-first prompting lets vLLM reuse the ~14K-token document prefix across queries via prefix caching.

Final references are fixed to Stage A; the abstractive answer comes from Stage B. A lighter single-model production variant (Qwen3-30B-A3B-FP8) is also containerized for deployment.

Infrastructure

• Serving: vLLM with FP8 / AWQ / NVFP4 quantization — both models fit one at a time on a single A100-40GB.
• Compute: the LANTA HPC cluster (SLURM, Lustre); inference dispatched as sbatch jobs.
• Packaging: reproducible Apptainer (Singularity) images for the benchmark's container backend.
• CI/CD: GitHub Actions + Google Cloud Build build and push images to a private registry.

Repository layout

The whole project is experiment-driven: every idea is a numbered expNN/ directory, scored on a held-out leak-free split, and tracked in SCORES.md.

Team

Built by a two-person team for the competition; I was the primary author of the pipeline, experiments, and serving infrastructure.