Configuration Gallery
Overview
| Aspect | Details |
|---|---|
| Purpose | Quick pointers to common presets and overlays. |
| Audience | Users looking for ready-to-use configurations. |
| Note | Presets are repo assets, not shipped in wheels. |
| Source | configs/presets/ and configs/overlays/. |
Pointers to common presets in this repository you can start from. Presets are
repo assets (not shipped in wheels). Use flag‑only invarlock evaluate when
installing from PyPI, or clone this repo to reference these files.
Note: Adapter‑based flows such as invarlock evaluate and invarlock run with
HF models require extras like invarlock[hf] or invarlock[adapters]. The
core install (pip install invarlock) remains torch‑free.
Presets (Runnable)
Causal LM (decoder-only)
| Preset | Use Case | Model Type | Dataset |
|---|---|---|---|
configs/presets/causal_lm/wikitext2_512.yaml | Standard evaluation | Decoder-only causal | WikiText-2 |
When to use: Primary preset for causal language models. 512-token sequences provide good coverage while keeping runtime reasonable.
invarlock evaluate --baseline gpt2 --subject /path/to/edited \
--preset configs/presets/causal_lm/wikitext2_512.yaml --profile ci
Masked LM (BERT, RoBERTa, etc.)
| Preset | Use Case | Model Type | Dataset |
|---|---|---|---|
configs/presets/masked_lm/wikitext2_128.yaml | Standard MLM evaluation | BERT/RoBERTa | WikiText-2 |
configs/presets/masked_lm/synthetic_128.yaml | Offline testing | BERT/RoBERTa | Synthetic |
When to use: MLM presets for BERT-family models. Use synthetic preset when network access is unavailable or for CI smoke tests.
invarlock evaluate --baseline bert-base-uncased --subject /path/to/edited \
--preset configs/presets/masked_lm/wikitext2_128.yaml --profile ci
Seq2Seq (T5, etc.)
| Preset | Use Case | Model Type | Dataset |
|---|---|---|---|
configs/presets/seq2seq/synth_64.yaml | Quick seq2seq tests | T5 | Synthetic |
When to use: Encoder-decoder models. Synthetic data keeps runs offline and fast for smoke testing.
Edit Overlays (Demo RTN Quantization)
These overlays apply the built-in quant_rtn edit for demonstration. For
production, use Compare & evaluate (BYOE) with your
own pre-edited checkpoint instead.
| Overlay | Scope | Use Case |
|---|---|---|
configs/overlays/edits/quant_rtn/8bit_attn.yaml | Attention layers only | Conservative quantization demo |
configs/overlays/edits/quant_rtn/8bit_full.yaml | All linear layers | Full model quantization demo |
configs/overlays/edits/quant_rtn/tiny_demo.yaml | Minimal layers | Quick smoke test |
Example (demo edit):
invarlock evaluate --baseline gpt2 --subject gpt2 \
--preset configs/presets/causal_lm/wikitext2_512.yaml \
--edit-config configs/overlays/edits/quant_rtn/8bit_attn.yaml \
--profile ci
Profiles
Profiles control window counts and bootstrap depth:
| Profile | Windows | Bootstrap | Use Case |
|---|---|---|---|
ci | 200/200 | 1200 | Standard CI evaluation |
release | 400/400 | 3200 | Production releases |
ci_cpu | 120/120 | 1200 | CPU-only environments |
Tips
- Use
--profile ci|release|ci_cputo apply runtime window counts and bootstrapping defaults. - Keep
seq_len = stridefor deterministic non‑overlapping windows. - Combine presets with edit overlays using multiple
-cflags or--edit-config. - For custom data, see Bring Your Own Data.
Related Documentation
- Configuration Schema — All config options
- CLI Reference — Command flags and profiles
- Compare & evaluate (BYOE) — Production workflow
- Dataset Providers — Available data sources