{"id":50820185,"url":"https://github.com/lappom/meetingqwen","last_synced_at":"2026-06-13T13:04:59.174Z","repository":{"id":362388709,"uuid":"1258848257","full_name":"Lappom/meetingqwen","owner":"Lappom","description":"LoRA fine-tuning pipeline to turn meeting transcripts into structured JSON — Qwen3.5-2B, Unsloth, GGUF/Ollama.","archived":false,"fork":false,"pushed_at":"2026-06-04T01:50:08.000Z","size":5700,"stargazers_count":0,"open_issues_count":0,"forks_count":0,"subscribers_count":0,"default_branch":"main","last_synced_at":"2026-06-04T03:11:11.981Z","etag":null,"topics":["fine-tuning","json","qwen"],"latest_commit_sha":null,"homepage":"","language":"Python","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":null,"status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/Lappom.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":null,"code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":null,"maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2026-06-04T01:23:50.000Z","updated_at":"2026-06-04T01:50:12.000Z","dependencies_parsed_at":null,"dependency_job_id":null,"html_url":"https://github.com/Lappom/meetingqwen","commit_stats":null,"previous_names":["lappom/meetingqwen"],"tags_count":null,"template":false,"template_full_name":null,"purl":"pkg:github/Lappom/meetingqwen","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/Lappom%2Fmeetingqwen","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/Lappom%2Fmeetingqwen/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/Lappom%2Fmeetingqwen/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/Lappom%2Fmeetingqwen/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/Lappom","download_url":"https://codeload.github.com/Lappom/meetingqwen/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/Lappom%2Fmeetingqwen/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":34285253,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-05-26T15:22:16.424Z","status":"online","status_checked_at":"2026-06-13T02:00:06.617Z","response_time":62,"last_error":null,"robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":true,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["fine-tuning","json","qwen"],"created_at":"2026-06-13T13:04:58.164Z","updated_at":"2026-06-13T13:04:59.169Z","avatar_url":"https://github.com/Lappom.png","language":"Python","funding_links":[],"categories":[],"sub_categories":[],"readme":"# MeetingQwen-2B\n\nFine-tuning pipeline for a lightweight meeting analysis model based on **Qwen3.5-2B** (`unsloth/Qwen3.5-2B`).\n\n**Input:** meeting transcript (optional speaker labels, timestamps)  \n**Output:** structured JSON (summary, decisions, action items, risks, open questions, next steps)\n\nTarget: local inference on consumer hardware, GGUF + Ollama compatible.\n\n## Pipeline\n\n```text\nAudio → Faster-Whisper → Transcript → MeetingQwen-2B → JSON → PDF / Markdown / Notion / Jira\n```\n\n## Project layout\n\n```text\nmeetingqwen/\n├── data/raw/          # Downloaded corpora\n├── data/generated/    # DeepSeek teacher labels\n├── data/cleaned/      # Validated samples\n├── data/train/        # train.jsonl + val.jsonl\n├── scripts/           # Data pipeline\n├── training/          # Unsloth train / merge / GGUF\n├── notebooks/         # Colab training notebook\n└── configs/training.yaml\n```\n\n## Setup\n\n```powershell\npython -m venv .venv\n.\\.venv\\Scripts\\Activate.ps1\npip install -r requirements.txt\ncopy .env.example .env\n# Edit .env — set DEEPSEEK_API_KEY\n```\n\n## Phase 1 — Download datasets\n\n```powershell\npython scripts/download_datasets.py\n```\n\nSources: QMSum, MeetingBank (Hugging Face). AMI requires manual registration — see `data/raw/ami_README.txt`.\n\n## Phase 2-3 — Generate teacher labels\n\n```powershell\n$env:DEEPSEEK_API_KEY = \"sk-...\"\npython scripts/generate_labels.py --limit 100   # test batch\npython scripts/generate_labels.py               # full run (target: 3000–25000)\n```\n\n## Phase 4 — Validate\n\n```powershell\npython scripts/validate_dataset.py\n```\n\n## Phase 5 — Build SFT dataset\n\n```powershell\npython scripts/build_sft_dataset.py\n```\n\nOnly `data/train/` is needed for fine-tuning (not `data/generated/` or `data/cleaned/`).\n\n### Dataset size targets\n\n| Level | Examples | Use case |\n|-------|----------|----------|\n| Pipeline test | 100–500 | Verify train → merge → GGUF workflow |\n| Minimum | 3,000 | First usable model |\n| Recommended | 10,000 | Good quality / cost balance |\n| Ideal | 25,000 | Best generalization |\n\n## Phase 6 — Fine-tune (GPU required)\n\nUnsloth requires an **NVIDIA GPU with CUDA**. It does not run on CPU or AMD GPUs (Windows).\n\n### Option A — Google Colab (recommended)\n\n1. Push the repo to GitHub with `data/train/` (use `git add -f` — see `.gitignore`).\n2. Open `notebooks/train_colab.ipynb` on Colab (Runtime → T4 GPU).\n3. Run all cells: train → merge → export GGUF → download `outputs/`.\n\n### Option B — Local CUDA machine\n\n```powershell\npip install \"unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git\"\npip install torch transformers trl accelerate bitsandbytes\npython training/train_unsloth.py\n```\n\nConfig: `configs/training.yaml` (LoRA r=16, 3 epochs, seq 32768).\n\n## Phase 7 — Evaluate\n\n```powershell\npython scripts/evaluate.py --limit 100\n```\n\nMetrics: ROUGE-L, BERTScore, JSON validity rate.\n\n## Phase 8 — Merge LoRA\n\n```powershell\npython training/merge_lora.py\n```\n\nOutput: `outputs/MeetingQwen-2B-merged/`\n\n## Phase 9 — GGUF export\n\n```powershell\npython training/export_gguf.py\npython training/export_gguf.py --ollama-create\n```\n\nQuantizations: Q4_K_M, Q5_K_M, Q8_0 → `outputs/gguf/`\n\nLocal inference with Ollama:\n\n```powershell\ncd outputs\\gguf\nollama create meetingqwen-2b -f Modelfile\nollama run meetingqwen-2b\n```\n\n## Phase 10 — Hugging Face release\n\nUpload merged weights, LoRA adapters, GGUF files, model card, and evaluation results to `MeetingQwen-2B`.\n\n## Output schema\n\n```json\n{\n  \"executive_summary\": \"\",\n  \"decisions\": [],\n  \"action_items\": [{ \"owner\": \"\", \"task\": \"\", \"deadline\": \"\" }],\n  \"risks\": [],\n  \"open_questions\": [],\n  \"next_steps\": []\n}\n```\n\n## Hardware targets\n\n- 8 GB RAM minimum\n- CPU inference possible (Q4_K_M GGUF via Ollama)\n- Fine-tuning: NVIDIA GPU (Colab T4 or local CUDA)\n- Consumer laptops, local-first\n\n## License\n\nBase model: follow Qwen3.5 license. Dataset sources: see respective corpus licenses.\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Flappom%2Fmeetingqwen","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Flappom%2Fmeetingqwen","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Flappom%2Fmeetingqwen/lists"}