constraint-carrier — direction & probe artifacts
Per-model, per-layer constraint direction vectors v̂_c (difference-in-means), extracted with the constraint-carrier harness.
Files
directions/Qwen__Qwen3-0.6B/<polarity>_<protocol>.safetensors # [L+1, D] unit directions + norms
directions/Qwen__Qwen3-1.7B/… # (same layout per model)
directions/Qwen__Qwen3-4B/…
directions/<model>/manifest.json # extraction config + provenance
⚠️ Read before use
- These directions are specific to one model, one layer, and one extraction protocol and do not transfer. In our pilot the last-token and max-pool directions were nearly orthogonal (median ∠ ≈ 88°). Always check which protocol you are loading.
- Activation reinjection with these vectors is a research tool, not a safety mechanism. Do not deploy it as a guardrail.
- Extracted from Qwen3 (Apache-2.0); this artifact inherits Apache-2.0. Each uploaded file lists its source model in
manifest.json. We do not claim any automatic license "carry-over" beyond what the upstream license grants.
How they were made
v̂_c^ℓ = normalize(mean_{with} h^ℓ_{t*} − mean_{without} h^ℓ_{t*}), over paired prompts that include vs omit the constraint (src/dirs.py in the harness repo). Reproduce from a clone of the harness:
python scripts/export_directions.py --model Qwen/Qwen3-0.6B # Apple MPS
python scripts/export_directions.py --model Qwen/Qwen3-4B --dtype bfloat16 # CUDA
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support