🧭 TinyLlama Activation Vector Steering

Model: TinyLlama/TinyLlama-1.1B-Chat-v1.0 + LoRA rank-16 adapter

Steering vectors are derived from BeaverTails:

v̂_layer = unit( μ_safe − μ_unsafe ) (layers 8–20)

How this Space applies steering (implementation):

• Where: Forward hooks on decoder layers 8–20; the direction is v̂ (same as above).
• Where in the sequence: The injection is applied only at the last token position (the position used for the next-token prediction at each step), not broadcast to the full sequence.
• Magnitude: Each step uses α as a relative strength: the update is proportional to ‖h‖ at that position and scaled by 1/√L over the L steered layers so activations stay stable (not raw α·v̂ in absolute units at every layer).
• Steering off: The Steering control is On/Off (off = steered column shows the same text as baseline — no second generation with hooks).
• Steered decoding: Temperature sampling (T=0.7, top-p=0.9) instead of greedy. Greedy + a fixed-direction hook injection creates a feedback loop that degenerates into repetition; sampling breaks it. Backstops: repetition penalty 1.35, no 6-gram repeats.
• Safety override: For clearly harmful prompts, the steered output is replaced with a safe-redirection response regardless of steering parameters.

⚠️ Running on CPU — each response can take ~30–90 seconds. Please be patient.

Layers (hooks): 8–20 · Hidden dim: 2048 · Adapter rank: 16 · v̂: unit(μ_safe − μ_unsafe)
Injection: last position only · Scale: α × ‖h‖ / √L (see steering.py) · Steered decode: T=0.7, top-p=0.9, rep. penalty 1.35, no 6-gram repeat