dialectic-dataset

Training pipeline for dialectic reasoning models — fine-tuning LLMs to engage with competing perspectives, identify tensions, and reason toward grounded conclusions.

Working Models

Model	Base	HuggingFace	Notes
8B v1	Qwen/Qwen3-8B	hikewa/dialectic-qwen3-8b-lora	Best overall
1.5B v1	Qwen/Qwen2.5-1.5B-Instruct	hikewa/dialectic-qwen2.5-1.5b-lora	Lightweight

Try them: HuggingFace Space

Project Structure

src/dialectic_dataset/    # Core library (scoring, formatting, training, generation)
scripts/                  # Pipeline scripts (generate, build, train, eval, upload)
tests/                    # Unit tests
data/                     # Training data, seeds, gold set
models/                   # Local LoRA adapters (not in git)
space/                    # HuggingFace Gradio Space
docs/                     # Specs, rubrics, post-mortem
archive/                  # GIGO artifacts from v2-v8 (not in git, do not reuse)

Pipeline

# 1. Generate traces
python scripts/generate_diverse_traces.py

# 2. Build training data (rejects fabrication, optional: keep thinking traces)
python scripts/build_training.py --traces data/traces.jsonl --output-dir data/training --keep-thinking

# 3. Train
python scripts/train.py --base-model Qwen/Qwen3-8B --training-dir data/training --output-dir models/my-lora

# 4. Evaluate
MISTRAL_API_KEY=... python scripts/eval.py --adapter-path models/my-lora/final

# 5. Dogfood (read the outputs yourself)
python scripts/dogfood.py

# 6. Upload
python scripts/upload.py --repo hikewa/my-model --adapter-dir models/my-lora/final

Lessons Learned

See docs/post_mortem.md for the full v1-v8 post-mortem. Key takeaway: train on thinking traces, not polished outputs. Dogfood after every version.

Tests

pip install -e ".[dev]"
pytest