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.
Integrity and exact runtime checks
- GGUF SHA-256:
d1494860fec450cbe61ee26ae17eba7723ee498c08001d88410e1aafe9bca8da
- Manifest SHA-256:
993bfbc702e906bb7547f7992adf81a02740e75b260e67cad23e648d936e8db6
- Ralph content hash:
133d377a000576179fe2a2ba85edd485bd7b2100df9769254a937d76bef91af2
- Declared compute: 3.5 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.
