DomainSteer

Contrastive activation addition for scientific domains. DomainSteer builds one steering vector per field and adds it to a causal language model's residual stream while the model generates.

Pairs for 144 scientific domains ship with the package, so building a vector needs no API key. The 2,160-item polysemy-trap benchmark ships too.

pip install domainsteer

from domainsteer import DomainSteerer

steerer = DomainSteerer(
    model_name="meta-llama/Llama-3.1-8B-Instruct",
    domain="Industrial biotechnology",
)
steerer.build(layer=15)
steerer.generate("What is a strain?", alpha=0.25)
steerer.compare(
    "If the assembly falls apart, do you get back the original pieces?",
    alpha=0.25,
)

Steered, "strain" is a microbial variant. Unsteered, it is a pulled muscle. Each bundled pair is that contrast: the same unnamed question, answered once in the domain sense and once in the everyday sense.

build(layer=15) extracts the vector at decoder block 15 and caches it. Later calls reuse the cache. generate takes a raw strength alpha (typical range 0.10–0.40). The shift at each token is h ← h + α‖h‖v.

Directions are written under ~/.cache/domainsteer/<domain>/<model>/directions/. Pass cache_dir= to put them somewhere else.

Bundled pairs

from domainsteer import list_bundled_models, bundled_manifest, load_bundled_pairs

list_bundled_models()
load_bundled_pairs("meta-llama/Llama-3.1-8B-Instruct", "3106")
load_bundled_pairs("meta-llama/Llama-3.1-8B-Instruct", "Industrial biotechnology")

Pairs are model-specific. Both shipped models share the question stems and do not share the answers. A vector built from 8B pairs is not applied to 3B. If the model you name has no bundled set, load_bundled_pairs returns None.

DomainSteerer looks for pairs in this order:

1. A cache file under cache_dir/<domain-slug>/pairs.jsonl
2. The bundled pairs for that model and domain
3. This model answering the 30 shared questions in data/questions/ itself, with the hook off. No API call. Used when the model is not one of the two shipped sets (for example Gemma).
4. API generation through PairGenerator (pip install "domainsteer[generation]" and an API key), only when the domain has no shipped questions

Turn the bundle off with use_bundled_pairs=False. Regenerate with build(force_pairs=True).

Benchmark

The polysemy-trap exam is 144 domains × 15 items = 2,160 questions. The question does not name the field. The overloaded term has to be read in the domain sense.

from domainsteer import load_benchmark, iter_benchmark

item = load_benchmark("3106")["items"][0]
item["question"]        # 'What can a fingerprint distinguish?'
item["gold"]            # 'product identity'
item["baseline_trap"]   # "a person's identity"
item["gold_responses"]  # one frozen reference sentence

sum(len(d["items"]) for d in iter_benchmark())  # 2160

gold_responses holds exactly one reference sentence per item.

What build does

1. Resolve the domain against the registry of 144 ANZSRC groups, ten concepts each.
2. Load the 30 contrastive pairs for that model and domain.
3. Replay both answers under the helper system prompt. The domain prompt is not in this forward pass.
4. Mean-pool the assistant-token hidden states at the layer you passed.
5. Save v = unit(mean expert − mean default).

Pass estimator="rfm" to use a Recursive Feature Machine instead of the difference of means. Those vectors are cached separately.

Calling build() with no layer asks an NLI judge to pick the layer and to map an expertise dial onto alpha. That judge is not installed by pip install domainsteer. Use build(layer=…) and a raw alpha.

Optional installs

License

MIT.