80 lines
3.3 KiB
Markdown
80 lines
3.3 KiB
Markdown
|
# Amygdala Readout Vector Training
|
|||
|
|
|
|||
|
|
Training pipeline that produces the safetensors file the vLLM
|
|||
|
|
ReadoutManager loads at runtime (see
|
|||
|
|
`vllm/vllm/v1/worker/readout_manager.py`). Produces per-hooked-layer
|
|||
|
|
`[n_concepts, hidden_size]` projection matrices keyed as
|
|||
|
|
`layer_<idx>.vectors` — the directions the runner projects residual
|
|||
|
|
activations onto during each forward pass.
|
|||
|
|
|
|||
|
|
## Overview
|
|||
|
|
|
|||
|
|
Two scripts, run in sequence:
|
|||
|
|
|
|||
|
|
1. **`extract_training_pairs.py`** — turns the memory graph into a
|
|||
|
|
directory of (emotion, polarity, text) training examples.
|
|||
|
|
Positive examples are memory nodes where the emotion scored
|
|||
|
|
≥ a threshold; negative examples are nodes where it's absent or
|
|||
|
|
low. Emotion tags come from the trailing `warmth:9 clarity:10 …`
|
|||
|
|
lines the subconscious agents emit.
|
|||
|
|
|
|||
|
|
2. **`train_steering_vectors.py`** — for each emotion, runs the
|
|||
|
|
target model over the positive and negative examples, captures
|
|||
|
|
residual-stream activations at the configured target layers, and
|
|||
|
|
computes `mean(positive) - mean(negative)` as the steering
|
|||
|
|
direction. Normalizes per-layer to unit length and saves the
|
|||
|
|
whole `[E, L, H]` matrix.
|
|||
|
|
|
|||
|
|
The output file is passed to vLLM via `VLLM_READOUT_VECTORS` together
|
|||
|
|
with a `VLLM_READOUT_MANIFEST` JSON listing concepts and hooked layer
|
|||
|
|
indices.
|
|||
|
|
|
|||
|
|
## Method
|
|||
|
|
|
|||
|
|
This is Contrastive Activation Addition (CAA, Rimsky et al.) applied
|
|||
|
|
to naturally-occurring emotion labels rather than hand-crafted
|
|||
|
|
contrast pairs. The shape of the signal we're recovering is "what
|
|||
|
|
direction in the residual stream corresponds to the model processing
|
|||
|
|
text-with-emotion-E vs. text-without". Because our training data was
|
|||
|
|
generated by the very model we're instrumenting (past-self's journal
|
|||
|
|
entries, digest nodes, pattern nodes), the signal should be unusually
|
|||
|
|
clean — the emotion labels and the text are already causally linked
|
|||
|
|
through a single model's forward pass.
|
|||
|
|
|
|||
|
|
## Usage (design — not yet runnable)
|
|||
|
|
|
|||
|
|
```
|
|||
|
|
# Step 1: memory graph → training data
|
|||
|
|
python -m training.amygdala_training.extract_training_pairs \
|
|||
|
|
--memory-mcp-url http://localhost:7777 \
|
|||
|
|
--output-dir /tmp/amygdala_training_data \
|
|||
|
|
--min-positive-score 8 \
|
|||
|
|
--max-negative-mentions 0 \
|
|||
|
|
--min-content-chars 40 \
|
|||
|
|
--max-examples-per-emotion 500
|
|||
|
|
|
|||
|
|
# Step 2: training data → steering vectors
|
|||
|
|
python -m training.amygdala_training.train_steering_vectors \
|
|||
|
|
--model Qwen/Qwen3.5-27B \
|
|||
|
|
--training-data-dir /tmp/amygdala_training_data \
|
|||
|
|
--target-layers 3,18,33,36 \
|
|||
|
|
--output /path/to/amygdala_vectors.safetensors \
|
|||
|
|
--dtype bf16 \
|
|||
|
|
--batch-size 4
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
## Open questions
|
|||
|
|
|
|||
|
|
- **Emotion selection**: enumerating which ~200 emotions to cover.
|
|||
|
|
Could be "most-common tags in the graph" (data-driven) or "from
|
|||
|
|
core-personality / pattern nodes" (human-curated). Probably both.
|
|||
|
|
- **Layer selection**: middle-to-late layers (~60–80% of depth)
|
|||
|
|
usually hold abstract semantic representations best; experiment
|
|||
|
|
with which layers give the cleanest linear separation per emotion.
|
|||
|
|
- **Cross-talk**: if two emotions are highly co-occurring (warmth +
|
|||
|
|
love, frustration + tiredness), their vectors will be close; that's
|
|||
|
|
fine as long as we don't pretend they're independent axes.
|
|||
|
|
- **Generalization**: vectors trained on our memory graph may not
|
|||
|
|
generalize to out-of-distribution text. Check by applying them to
|
|||
|
|
held-out conversation data and eyeballing the projections.
|