Giorgio Gilestro
e433e48860
Adds the "directed sex" operator (E10) the moe regime-flip pointed to: don't
commit to one a-priori blend — breed a population of recombinant offspring
(specialists merged at Dirichlet-sampled weights), score each on a held-out
validation split with the verifier, and keep the fittest, reported on a fresh
test split. Two breeding objectives: best-overall and best-worst-family.
src/llm/directed.py + kind llm_directed, reusing the cached specialists.
Result — refinements pay off in proportion to how far the uniform soup is from
optimal:
- 0.5B (soup dilutes): directed selection beats soup on the bred objective —
directed_overall 0.69 > soup 0.64; directed_balanced worst-family 0.37 > 0.26.
Riders: single-objective selection trades off the other axis (overall-breed
tanks lists to 0.17); a global blend still trails per-input routing (0.74).
- 7B (Imperial CX3, soup already composes to ceiling on near-saturated families,
strings/arith 1.00): directed ~= soup (0.868 ~ 0.873, marginally below via a
val/test overfit gap) — no fitter offspring to breed.
Through-line across all four LLM runs: "merge, don't average" and its refinements
(routing, directed selection) are weak-base / suboptimal-default phenomena — they
help at 0.5B and are inert at 7B. Honest limitation kept in the writeup: the 7B
families are near-saturated, which caps the headroom; a harder unsaturated
benchmark is the fair next test.
Also folds in the two llm_moe local manifest/config files missed in 8da0dac.
+3 directed unit tests (130 green). Results in results/llm_directed{,_hpc}/
(parquet gitignored).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
25 lines
1.3 KiB
YAML
25 lines
1.3 KiB
YAML
experiment: llm_directed
|
|
kind: llm_directed
|
|
seed: 1
|
|
n_replicates: 1
|
|
|
|
# Layer 2 / LLM — DIRECTED SEX (E10) in weight space: breed many recombinant offspring and keep the
|
|
# fittest. llm_moe showed fusion can COMPOSE beyond the parents (7B) but the right blend is unknown and
|
|
# base-dependent, while pure routing is capped at the best parent. Directed sex resolves both: sample a
|
|
# population of weighted merges of the specialists, score each on a held-out VALIDATION split with the
|
|
# verifier (grounding), and select the winner — reported on a fresh TEST split (no selection-on-test
|
|
# leakage). Two objectives: best-overall and best-worst-family (raw capability vs the balanced
|
|
# generalist). Prediction: directed selection beats the single a-priori soup and every parent, at
|
|
# either scale. Falsifier: directed offspring ≤ uniform soup on test.
|
|
|
|
base_model: Qwen/Qwen2.5-0.5B-Instruct # reuses the cached llm_merge specialists
|
|
families: [lists, strings, arith]
|
|
n_train: 700 # only if cached specialists are absent
|
|
n_val: 80 # held-out split the verifier selects on
|
|
n_test: 100 # fresh split winners are reported on
|
|
n_candidates: 16 # offspring population size
|
|
concentration: 0.5 # Dirichlet concentration (<1 = sparser, specialist-dominant blends)
|
|
epochs: 3
|
|
lora: {r: 16, alpha: 32}
|
|
|
|
output: {dir: results/llm_directed}
|