Qwen2.5-0.5B-RYS-3-7-GGUF

A layer-duplication ("RYS" โ€” Repeat Your Self, David Ng) variant of Qwen/Qwen2.5-0.5B-Instruct: layers 3โ€“6 duplicated, 24 โ†’ 28 layers. No training, merging, or weight changes. GGUF (imatrix Q-quants).

โš ๏ธ Evaluation status โ€” please read (updated 2026-06)

On a controlled re-test, the headline gains for this model do not hold up. This card is being corrected to say so.

The original card reported reasoning 35.29% โ†’ 58.82% (+23.53%) and an EQ lift (37.38 โ†’ 55.23), calling this an "EQ specialist." A re-run on a current llama.cpp build found:

  • Reasoning does not reproduce. The baseline reproduces exactly (35.29%), but the (3,7) config scored 29.41% โ€” below the baseline. The original +23.5% was probe variance (the reasoning probe is 17 questions, so each one is ยฑ5.9%).
  • The EQ number is not a reliable capability signal. On output inspection, the EQ probe on this model is satisfied by emitting a near-constant rating vector across different scenarios โ€” it outscores the base model's degenerate output without actually reading the scenario. So the "EQ lift" reflects the scoring metric, not emotional understanding.

These scores come from a lightweight search probe (16 math / 16 EQ / 17 reasoning questions) used to find productive layer blocks, not a validated benchmark. A standard-harness re-eval (lm-eval-harness / EQ-Bench at full N) is in progress.

Bottom line: treat this as a normal Qwen2.5-0.5B-Instruct with layers 3โ€“6 duplicated. Published for transparency of the RYS sweep โ€” not as an established "EQ specialist."

Original sweep numbers (search probe โ€” kept for the record)

probe reported baseline reported (3,7) re-test note
Reasoning (17 q) 35.29% 58.82% did not reproduce โ€” (3,7) โ†’ 29.41% (below baseline)
EQ (16 q) 37.38 55.23 EQ probe gamed by a constant guess; not validated EQ
Math (16 q) 0.492 0.424 search-probe score

Run it

llama-server -m Qwen2.5-0.5B-RYS-3-7-Q4_K_M.gguf -ngl 99

Method ยท data ยท attribution

  • Method: layer duplication โ€” 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 these probe deltas as capability is what this update corrects.

License

Apache-2.0.

Downloads last month
23
GGUF
Model size
0.6B params
Architecture
qwen2
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for john-broadway/Qwen2.5-0.5B-RYS-3-7-GGUF

Quantized
(395)
this model

Collection including john-broadway/Qwen2.5-0.5B-RYS-3-7-GGUF