Giorgio Gilestro
84124de143
Clarity pass over the main text (36-item audit), Discussion rewrite and cut, acknowledgements, Souly et al. as ref 62, lettered SI panels, model section moved under Results; plus the untracked curriculum/society/compose/smol configs, runners, figures, stats and tests that the SI already cites. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Y64o8FKP7rCuXzC48pxpMm
1.4 KiB
llm_merge_seeds_smol — the Fisher–Muller replication on a second base lineage (SmolLM2-1.7B-Instruct)
Claim tested (manuscript revision 2026-09-12). Every language-model result in the paper used one
model lineage (Qwen2.5). Does "merged specialists exceed every parent" hold on an unrelated base?
Base: HuggingFaceTB/SmolLM2-1.7B-Instruct (Apache-2.0, Llama architecture, different lab and
pretraining corpus). Identical protocol to llm_merge_seeds (three disjoint families, 600 training
tasks, 3 epochs, rank-16 LoRA, fixed 100-task test sets, soup and TIES merges, 5 training seeds);
adapters_dir: models/llm_smol keeps the specialist cache apart from Qwen's. Ran on one L40S in
12 min (hpc/llm_smol.pbs). Numbers from figures/stats_llm_smol.py; figure llm_smol.png (both
lineages side by side).
Result — replicates, 5/5 seeds, with the same magnitudes
| contrast (SmolLM2) | overall | worst family |
|---|---|---|
| soup − best specialist | +0.049 ± 0.022 (5/5 seeds) | +0.190 ± 0.045 (5/5) |
| TIES − best specialist | +0.097 ± 0.020 (5/5) | +0.202 ± 0.077 (5/5) |
| (Qwen2.5-0.5B, for reference) soup − best | +0.040 ± 0.035 (4/5) | +0.200 ± 0.020 (5/5) |
| (Qwen) TIES − best | +0.055 ± 0.020 (5/5) | +0.204 ± 0.016 (5/5) |
Absolute levels: best specialist 0.610 ± 0.013 overall / 0.128 worst-family; soup 0.659 / 0.318; TIES 0.707 / 0.330. Falsifier (not triggered): merged ≤ best specialist in a majority of seeds.