dancinlab/composite-v1-dev-r6-multi

composite-v1-lora-r6-multi

One-line summary: Round-6 multi-substrate LoRA adapter set (4 substrates) — φ★ 14-gate substrate study.

• Family: composite
• Stage: lora-r6
• Step: final
• Substrate: 4-way (Qwen3-8B · Qwen2.5-7B · Mistral-Nemo-12B · Gemma-3-12B)

Origin

Round-6 of the multi-substrate LoRA training series (trained_adapters_r4/r5/r6). Four sibling adapters, one per substrate: p1=Qwen3-8B (r64), p2=Qwen2.5-7B (r64), p3=Mistral-Nemo-12B (r96), p4=Gemma-3-12B (r128). Part of the φ★ (phi-star) 14-gate substrate study (LORA.md). r4/r5/r6 are distinct rounds (sha256-verified non-identical).

Falsifiers

• Substrate-invariance claim FALSE if the 4 adapters' behaviors diverge beyond the study's tolerance on the shared evaluation (per the φ★ gate criteria).
• "round-6 improves on r5" FALSE if r6 does not beat r5 on the gate metrics.

Substrate

Four base models adapted via LoRA (ranks 64/64/96/128). No merged weights — adapters only; load atop the respective base. training_args.bin included per adapter.

Caveats

• WIP / intermediate research adapters — PRIVATE, not released models.
• Adapter-only: requires the matching base model (not redistributed here).
• Verification per φ★ gates; loss/perplexity are not truth (p7). Local-only backup → HF.

Composability

Series: trained_adapters_r4 → r5 → r6 (this) → partial r7 (p4 only). Each adapter composes atop its base substrate. Part of the composite LM family.