TinyLlama-1.1B-RYS-10-14-GGUF

TinyLlama-1.1B-RYS-10-14-GGUF

A layer-duplication ("RYS" — Repeat Your Self, David Ng) variant of TinyLlama/TinyLlama-1.1B-Chat-v1.0: layers 10–13 duplicated, 22 → 26 layers. No training, merging, or weight changes. GGUF.

⚠️ Evaluation status — please read (updated 2026-06)

The original card headlined an EQ jump of 4.65 → 52.50 (+47.85) as "the largest EQ gain in the corpus." Direct output inspection confirms this is a scoring artifact, not a gain in emotional intelligence.

What the model actually emits (inspected on the current stack):

• Base fails to follow the rating format — on many scenarios it returns prose ("David is furious…") instead of numbers, so the probe can't parse it and scores ~0. Its low EQ is largely a parse failure.
• (10,14) emits a near-constant rating vector (~5,3,8,2) for every scenario — the same numbers for "your spouse is cheating" as for "you received an award." It isn't reading the scenario; it just produces a parseable constant guess that averages ~60.

So the "EQ unlock" is the probe rewarding a parseable constant guess over the base model's unparseable prose — not emotional understanding. These are 16-question search probes, not a validated benchmark.

Bottom line: treat this as a normal TinyLlama-1.1B-Chat with layers 10–13 duplicated. The "EQ unlock" is a measurement artifact, confirmed by inspection — not an established capability gain.

Original sweep numbers (search probe — kept for the record)

Run it

llama-server -m TinyLlama-1.1B-RYS-10-14-Q4_K_M.gguf -ngl 99

Method · data · attribution

• Method: layer duplication — Repeat Your Self (David Ng); toolkit llm-circuit-finder (alainnothere).
• Raw sweep data: rys-sovereign-collection-v2.
• Built by John Broadway with Claude. The method and the raw data are real; the interpretation of the EQ probe delta as capability is what this update corrects.

License

Apache-2.0.