README.md

metadata

title: DOTA2Tuned
sdk: gradio
app_file: app.py
python_version: '3.12'
tags:
  - track:backyard
  - track:wood
  - sponsor:openbmb
  - sponsor:openai
  - sponsor:modal
  - achievement:welltuned
  - achievement:offbrand
  - achievement:sharing
models:
  - build-small-hackathon/dota2tuned-qwen3-4b-2507-lora
  - build-small-hackathon/dota2tuned-minicpm4-1-8b-lora
  - build-small-hackathon/dota2tuned-qwen3-30b-a3b-2507-lora
datasets:
  - build-small-hackathon/dota2tuned-data

DOTA2Tuned

DOTA2Tuned is a Hugging Face Build Small Hackathon project for Dota 2 drafting, meta analysis, build suggestions, and match prediction.

The implementation is designed around a simple rule: stats and predictors choose the recommendations, while a small fine-tuned model explains them with patch-aware evidence.

• Hugging Face Space: https://build-small-hackathon-dota2tuned.hf.space
• Modal alternate UI: https://dracufeuer--dota2tuned-ui.modal.run
• Tiny fine-tuned adapter: https://huggingface.co/build-small-hackathon/dota2tuned-qwen3-4b-2507-lora
• Balanced fine-tuned adapter: https://huggingface.co/build-small-hackathon/dota2tuned-minicpm4-1-8b-lora
• Quality fine-tuned adapter: https://huggingface.co/build-small-hackathon/dota2tuned-qwen3-30b-a3b-2507-lora
• Dataset artifacts: https://huggingface.co/datasets/build-small-hackathon/dota2tuned-data
• Demo video: TODO add final demo video URL before validation.
• Social post: TODO add final social post URL before validation.

Hackathon Validation

The Build Small validator checks this Space README. Entry requirements are: sub-32B models, Gradio Space in the Build Small org, demo video, social-media post linked from this README, and GPU-limit compliance.

Validator: https://build-small-hackathon-field-guide.hf.space/submit

Selected tracks, prizes, and badges:

• Tracks: Backyard AI, Thousand Token Wood.
• Sponsor prizes: OpenBMB Best MiniCPM Build, OpenAI Best Use of Codex, Modal Best Use of Modal.
• Bonus badges: Well-Tuned, Off-Brand, Sharing is Caring.
• Not claimed: Nemotron Hardware Prize, Off the Grid, Llama Champion.
• GPU note: this Space does not use Zero GPU allocation; training and tuned-model serving use Modal.

Social post draft:

Built DOTA2Tuned for the Hugging Face Build Small Hackathon: a Gradio Dota 2 draft coach that combines STRATZ/OpenDota match evidence, deterministic draft stats, and fine-tuned sub-32B adapters for grounded explanations.

It suggests heroes, counters, synergies, builds, match predictions, and caveats weak data instead of inventing unsupported meta claims.

Space: https://build-small-hackathon-dota2tuned.hf.space Tiny: https://huggingface.co/build-small-hackathon/dota2tuned-qwen3-4b-2507-lora Balanced: https://huggingface.co/build-small-hackathon/dota2tuned-minicpm4-1-8b-lora Quality: https://huggingface.co/build-small-hackathon/dota2tuned-qwen3-30b-a3b-2507-lora Repo: https://github.com/1ncompleteness/DOTA2Tuned

Quick Start

uv sync
cp .env.example .env
uv run dota2tuned health
uv run dota2tuned smoke --live
uv run dota2tuned ingest --pro-matches 100 --public-matches 100 --enrich-limit 20
uv run dota2tuned normalize
uv run dota2tuned features
uv run dota2tuned build-rag
uv run dota2tuned serve

Update .env with your Hugging Face, STRATZ, OpenDota, and Steam tokens before running large ingestion or Hub operations. Fine-tuning now runs on Modal by default for this project. Set MODAL_ENABLED=1, MODAL_TOKEN_ID, and MODAL_TOKEN_SECRET, then use uv sync --extra modal before deploying Modal functions. HF_TOKEN must still include repo.write so the training run can push adapters to the configured Hub model repos. Use MODEL_PROFILE or --profile to select Tiny, Balanced, or Quality.

The submitted Space includes compact serving artifacts under data/parquet, data/rag, and data/models. Raw API responses remain local-only and ignored by git.

Main Commands

• dota2tuned ingest fetches raw reference data, pro matches, public matches, patch notes, and optional match enrichment.
• dota2tuned smoke --live validates configured tokens with tiny live API checks.
• dota2tuned normalize converts raw JSONL into Parquet tables and refreshes DuckDB views.
• dota2tuned features refreshes DuckDB views and reports feature table row counts.
• dota2tuned train-predictor trains the draft win predictor from normalized matches.
• dota2tuned build-rag creates patch/stat cards and a local retrieval index.
• dota2tuned make-sft creates JSONL examples for SFT.
• dota2tuned modal-deploy deploys the Modal Gradio app and GPU training function.
• dota2tuned modal-smoke validates the deployed Modal app can load artifacts.
• dota2tuned modal-train --profile qwen3_4b_2507 submits the Tiny Modal GPU QLoRA run. Other profiles include minicpm4_1_8b and qwen3_30b_a3b_2507; Quality routes to the H200-backed train_sft_quality function.
• uv run python scripts/watch_modal_training.py --call tiny=fc-... --interval 300 polls detached Modal training calls without blocking local development.
• dota2tuned modal-ask calls the fine-tuned adapter through Modal GPU inference.
• dota2tuned finetune --launch-job remains available for Hugging Face Jobs if a token has job.write.
• dota2tuned serve launches the Gradio app.

The Gradio app includes searchable hero, item, role, scope, and ability dropdowns, alias-aware hero lookup (PA, CM, AM, QOP, KOTL, and curated initials), selected icon previews, and six sidebar modules: Ask, Draft, Meta, Builds, Predictor, and Data. Ask calls the Modal generate_answer function when Modal credentials are configured in the runtime environment; otherwise it degrades with a clear unavailable message.

See PLAN.md for the full architecture and delivery plan. See MODEL_SELECTION.md for the current LLM decision and eval protocol. See MODEL_CARD.md for the model card mirrored to the Hub adapter repo. See DATASET_CARD.md for the dataset card mirrored to the Hub dataset. See SUBMISSION.md for the final hackathon handoff, demo script, social post draft, and verification checklist.