seb-vil/llama-3.2-3b-stoic-steering

Llama-3.2-3B Stoic Steering — frozen research artifacts

Frozen reference artifacts for the Stoic Steering project (GitHub): steering Stoic philosophical reasoning into meta-llama/Llama-3.2-3B via contrastive activation addition (CAA) and LoRA, with mechanistic interpretability checks.

Core finding: under matched-decoding measurement, CAA at the canonical coefficient moves nothing measurable (style, judge-scored content, or decisions), while LoRA genuinely shifts forced-choice decisions (Seneca ΔP(stoic) +0.061, both stance buckets; judge-free logit instrument).

These files are the frozen regression fixtures the repo's verification stages check against — they are inputs to reproduction, not a packaged steering product.

Contents

Path	What it is	Used by
steering_vectors/{author}_steering_3B.pt	CAA steering vectors, {layer: tensor} dicts (Marcus L26, Seneca L4, Epictetus L8)	Stage 2: fresh extraction must hit cosine ≥ 0.99 against these
lora_{author}_clean/	LoRA adapters (r=8, α=32, q_proj+v_proj, 3 epochs on author corpora)	Stage 4: merged onto a fresh base for the judge-free dilemma eval

Tokenizer: load from the base model, not from the adapter folders.

Fetch + verify

From a checkout of the GitHub repo:

python scripts/fetch_artifacts.py   # downloads into place, verifies checksums

SHA256 (also pinned in the repo's data/MANIFEST.sha256):

dc9d9512d5f5f25f3374cb2c976e4a101ac852e139bc98024dff56eee2926b2b  steering_vectors/epictetus_steering_3B.pt
f4b3c0589a6df40f2e70bed2468688143d961e7a9e8f247bf33198b89ac52f0c  steering_vectors/marcus_aurelius_steering_3B.pt
273d5fd739bda0cb3ad4814da2d485999c340c87bf448d843f0b3c9195c7ae6b  steering_vectors/seneca_steering_3B.pt
979000e33a57a85bf5ffd6078783ab0520750a3d4b992a6b971fa4d3eeb0d0cc  lora_epictetus_clean/adapter_model.safetensors
81d3a0e38ea69ff436d7454471693586b888f769a9d86427d23963876454e8bd  lora_marcus_clean/adapter_model.safetensors
c315155006e9d664012fc09cac51a3cc45655676f221d9ceaccf1b99b0bd83f9  lora_seneca_clean/adapter_model.safetensors

License

The adapters and steering vectors are derivatives of Llama 3.2 and are distributed under the Llama 3.2 Community License. The project's own code is MIT (see the GitHub repo).