README.md

---
tags:
- qwen3_5
- rys
- relayer
- safetensors
base_model: Jackrong/Qwopus3.6-27B-v2
---

# Qwopus3.6-27B-v2-RYS-Balanced

This is a RYS/relayer export of [`Jackrong/Qwopus3.6-27B-v2`](https://huggingface.co/Jackrong/Qwopus3.6-27B-v2).

The checkpoint physically duplicates selected decoder layers in the Hugging Face `safetensors` weights. It does not require the RYS runtime wrapper.

## Variant

- Target repo: `hampsonw/Qwopus3.6-27B-v2-RYS-Balanced`
- Objective: Balanced Math+EQ
- Repeated block: `15,30` (`--blocks "15,30"`)
- Repeated source layers: 15–29 inclusive, zero-indexed
- Source text layers: 64
- Target text layers: 79
- Extra repeated layers: 15

## Probe result

From the BF16 Transformers scan over `math_16 + eq_16`:

| Metric | Score | Delta vs baseline |
|---|---:|---:|
| Math | 0.818086 | +0.104831 |
| EQ | 0.741500 | +0.010000 |

Rank note: balanced rank 1.

Baseline scores from the scan:

- Math: `0.713255`
- EQ: `0.731500`


## Local diagnostic evaluation results

These are **local diagnostic subsample results**, not full benchmark claims. They were run through an OpenAI-compatible vLLM endpoint with Qwen3.6 thinking-mode sampling:

- `temperature=1.0`, `top_p=0.95`, `top_k=20`, `min_p=0.0`
- `thinking_token_budget=32768`
- `max_tokens=81920`
- MATH prompt: `Please reason step by step, and put your final answer within \boxed{}.`

Comparisons below use:

- Qwen baseline: `cyankiwi/Qwen3.6-27B-AWQ-BF16-INT4`
- Qwopus baseline: `Jackrong/Qwopus3.6-27B-v2`
- This model: `hampsonw/Qwopus3.6-27B-v2-RYS-Balanced`

### MATH-500 hardest-50 diagnostic

All three models solved the 50-item slice after manual normalization of mathematically equivalent answer formats. Raw scorer misses were formatting-equivalence issues such as `5.5` vs `\frac{11}{2}` and `2\sqrt{3}+1` vs `1+2\sqrt{3}`.

| Model | Raw scorer | Audited | Completion tokens | Reasoning tokens | Total tokens |
|---|---:|---:|---:|---:|---:|
| Qwen3.6-27B baseline | 47/50 | 50/50 | 633,524 | 599,909 | 639,736 |
| Qwopus3.6-27B-v2 | 46/50 | 50/50 | 304,733 | 270,520 | 310,945 |
| Qwopus3.6-27B-v2-RYS-Balanced | 45/50 | 50/50 | 301,726 | 267,582 | 307,938 |

On this MATH diagnostic slice, RYS-Balanced matched Qwopus v2 audited accuracy while using roughly the same number of tokens, and both Qwopus variants used about half the completion tokens of the Qwen baseline.

### LiveCodeBench release_v6 hardest-49 diagnostic

This diagnostic uses the deterministic hardest-50 slice from public `livecodebench/code_generation_lite release_v6`, with item `3344` dropped as a whole question because it contains a malformed private testcase outside the stated `List[List[int]]` function contract.

| Model | Correct | Accuracy | Completion tokens | Reasoning tokens | Total tokens |
|---|---:|---:|---:|---:|---:|
| Qwen3.6-27B baseline | 43/49 | 87.8% | 1,640,947 | 1,440,065 | 1,672,212 |
| Qwopus3.6-27B-v2 | 45/49 | 91.8% | 1,400,939 | 1,347,042 | 1,432,204 |
| Qwopus3.6-27B-v2-RYS-Balanced | 32/49 | 65.3% | 983,974 | 806,168 | 1,015,803 |

RYS-Balanced was substantially cheaper in tokens on this LCB diagnostic, but accuracy was much worse than both baselines. It had two terminal failed items (`2952`, `3233`) and many more non-passing solutions.

### Interpretation

This checkpoint is currently best viewed as a **math-focused experimental RYS export**. The small MATH diagnostic looked strong and token-efficient, but the LiveCodeBench diagnostic regressed significantly. Use with caution for coding-heavy workloads.

## Provenance

- Source model: `Jackrong/Qwopus3.6-27B-v2`
- Export method: RYS physical layer duplication
- Export manifest: `rys_export_manifest.json`
- Probe result bundle: `qwopus36-bf16-results_20260523T213305Z.tar.zst`

## Notes

This model has not yet been validated on the larger `math_120 + eq_140` probe set or full public benchmark suite. The local diagnostics above suggest the RYS-Balanced export may preserve math performance while reducing tokens, but it regressed on the coding diagnostic. Treat it as experimental.