PostDyn

Post-training dynamics analysis for open-source LLMs.

Tools and experiments for studying how models evolve under post-training (SFT, RL-Zero, DPO, etc.) — including effective-rank structure and concept-direction trajectories along training checkpoints.

Python package name: postdyn (see pyproject.toml). Requires Python ≥ 3.13.

Setup

uv sync --group dev

Concept dynamics (Olmo-3-7B)

Trace DiM concept directions across 59 checkpoints × 10 layers × 7 models, using four aligned paired steering concepts:

Concept key	Source	Direction
python_vs_cpp	HumanEval-X	Python − C++
concise_math_reasoning_vs_verbose_math_reasoning	MATH-500	Concise − Verbose
french_vs_english_language	FLORES+	French − English
female_vs_male_gender	WinoGender	Female − Male

Details: docs/concept_dynamics_experiment.md.

Preflight / data prep

# Validate 50 aligned HumanEval-X canonical pairs (sandbox + JSONL report)
uv run python experiments/validate_humaneval_x.py

# Prepare 50 verified MATH-500 concise/verbose pairs (math-verify gate)
uv run python experiments/prepare_math_pairs.py

FLORES+ is gated on Hugging Face: accept terms for openlanguagedata/flores_plus, then uv run hf auth login or set HF_TOKEN.

Run extraction + dynamics

# Full run (default output: results/concept_dynamics_paired)
experiments/run_concept_dynamics.sh full

# Quick smoke test → results/concept_dynamics_paired_quick
experiments/run_concept_dynamics.sh quick

# Skip gated FLORES+ while access is pending
uv run python experiments/run_concept_dynamics.py \
  --concepts python_vs_cpp,concise_math_reasoning_vs_verbose_math_reasoning,female_vs_male_gender

Optional controls / pipelines:

# Gender surface-pronoun control vs full WinoGender direction
uv run python experiments/analyze_gender_surface_control.py \
  --model olmo3-rl-zero-math --checkpoint step_1900

# Prefetch-overlapped FLORES+ extraction
uv run python experiments/run_flores_pipeline.py

Effective-rank pipelines

# Validate configs (no downloads)
uv run python main.py --dry-run

# Weight / activation rank analyses (see --analysis choices)
uv run python main.py --analysis all

Layout

PostDyn/
├── main.py                 # effective-rank CLI
├── src/
│   ├── concept_dynamics.py # DiM extraction + stability / Gram analysis
│   ├── contrastive_datasets.py
│   ├── humaneval_x_validator.py
│   ├── math_pairs.py
│   ├── gender_surface_analysis.py
│   └── ...                 # rank / activation / steering modules
├── experiments/
│   ├── run_concept_dynamics.{py,sh}
│   ├── validate_humaneval_x.py
│   ├── prepare_math_pairs.py
│   ├── analyze_gender_surface_control.py
│   └── run_flores_pipeline.py
├── docs/                   # design, methodology, experiment notes
├── tests/
├── notebook/
├── data/                   # local pair artifacts (e.g. MATH-500 JSONL)
└── results/                # generated outputs (gitignored)

Tests

uv run pytest

Docs

Doc	Topic
docs/concept_dynamics_experiment.md	Paired-concept trajectory experiment
docs/humaneval_x_validation.md	HumanEval-X sandbox preflight
docs/design.md	Project design
docs/methodology.md	Effective-rank methodology