boweizh1204/ralph-qwen3-8b-sub2-championship-final

Qwen3-8B Ralph v2 sub2 — championship hybrid v1

Miner-produced sub2 candidate for Ralph v2 on Bittensor netuid 40. This repository contains one behavior-bearing GGUF plus its validator-compatible manifest. It is isolated from the author's pre-existing sub4 artifact.

Immutable parent and training lineage

• Parent: Qwen/Qwen3-8B@b968826d9c46dd6066d109eabc6255188de91218
• Architecture: Qwen3, 36 layers, hidden size 4096, vocabulary size 151936
• Validator-measured weight elements: 8,190,427,136
• Training: rank-32 QLoRA, alpha 64, dropout 0.05, one epoch, effective batch 16
• Targets: attention q/k/v/o and MLP gate/up/down projections
• Merge point: update 27 of 54 (the half checkpoint outperformed the final checkpoint after low-bit quantization)
• Validation loss: 0.583881 before training; 0.100861 at the final adapter

The training set contains 864 custom parent-distilled examples, with a separate 96-example validation split. It is balanced across English, Hindi, and Chinese; shallow and deep reasoning; and math, code, dialogue, and agentic tasks. Published Ralph scored-pool items were used only as held-out diagnostics and were not used for training or calibration.

Importance calibration and quantization

The importance matrix was produced from 144 custom distilled examples: exactly six per language × depth × task cell, totaling 70,018 tokens and 34 chunks of 2,048 tokens.

• Attention projection matrices: Q2_K
• Feed-forward matrices and output head: true llama.cpp IQ2_S
• Token embedding: Q2_K
• Logical code bits/weight: 2.0000
• Physical container bits/weight: 2.5625
• Serialized package size: 2,647,344,636 bytes
• Tensor histogram: IQ2_S=6,058,147,840, Q2_K=2,132,279,296, F32=308,224
• Four-chunk calibration perplexity: 2.7210

Alternative pure IQ2_S, IQ2_XS, IQ2_XXS, TQ2_0, and importance-mixed layouts were screened. The attention-Q2_K hybrid produced the strongest signed-round Phi diagnostic while remaining inside both current sub2 limits.

The tuned all-IQ2_S finalist was then compared directly with the hybrid on both signed draws under every observer. Combined worst slices were hybrid 0.19645 vs all-IQ2_S 0.17363 on Smol, 0.21441 vs 0.21537 on Phi, and 0.21619 vs 0.23602 on OLMo. Cross-observer worst-case selection therefore favors the hybrid; all-IQ2_S also showed a confident Smol English/deep regression on signed round 2.

Integrity and exact runtime checks

• GGUF SHA-256: d1494860fec450cbe61ee26ae17eba7723ee498c08001d88410e1aafe9bca8da
• Manifest SHA-256: 9223ffb462371deafcbac08b922eca7146f9e549c627a5c2f1f37d41096581e6
• Ralph content hash: 9d71cedb7443b378c1619883f323c68e099812885c448c82a68341ae3ccd522d
• Declared compute: 4.3 H100-hours

The exact two-file package passes Ralph v2 inspect_checkpoint, the sub2 code/container bit gate, and pinned-parent compatibility. It test-loads and generates under llama-cpp-python==0.3.34 with CUDA, n_ctx=8192, n_batch=512, and all layers offloaded. On each of two signed post-commit selections, three independent runs (one per observer) produced byte-identical hashes for all 72 generated steps.

Held-out observer diagnostics

The fast immutable baseline and this candidate were scored on the same two actual signed post-commit selections (144 samples per observer). Frozen parent steps came from signature-verified public records. When an observer matched a signed record, the published continuation was reused exactly; otherwise a new continuation was generated. All distributions and candidate effects were recomputed locally.

Observer	Fast worst slice	Championship worst slice	Raw delta	Paired 95% LCB
SmolLM2-1.7B-Instruct	0.17706	0.19645	+0.01939	+0.00256
Phi-3-mini-4k-instruct	0.21511	0.21441	-0.00070	-0.01444
OLMo-2-1124-7B-Instruct	0.21434	0.21619	+0.00185	-0.01981

No observer showed a statistically confident regression axis. The candidate was materially better on Smol and essentially tied across the two draws on Phi and OLMo. These paired lower bounds do not clear Ralph's +0.05 contested-crown margin, so they must not be interpreted as a robust dethrone claim.

The candidate passed the formal repetition gate on both signed selections: 18/72 and 14/72 outputs exceeded the six-gram loop threshold. The fast baseline failed the first selection at 22/72 (30.56%) and passed the second at 21/72 (29.17%). Every observer run scored 72/72 shared samples with zero discards and no scorer reasons.

These are unsigned miner-side held-out diagnostics, not validator verdicts or crown certificates. At the last pre-upload status refresh, the sub2 lane had no commitment and no incumbent crown.