ralph-qwen3-8b-sub2

Qwen3-8B Ralph v2 sub2 — all-IQ2_S v3

Miner-produced sub2 candidate for Ralph v2 on Bittensor netuid 40. This repository contains a single behavior-bearing GGUF plus its validator-compatible manifest.

Immutable parent

• Model: Qwen/Qwen3-8B
• Revision: b968826d9c46dd6066d109eabc6255188de91218
• Architecture: Qwen3, 36 layers, hidden size 4096, vocabulary size 151936
• Validator-measured weight elements: 8,190,427,136

The BF16 parent was converted without an intermediate requantization. An importance matrix was built from a custom multilingual calibration set balanced across English, Hindi, and Chinese; shallow and deep prompts; and math, code, dialogue, and agentic tasks. No published Ralph scored-pool item was used for calibration or training.

Quantization

• Matrix weights: true llama.cpp IQ2_S
• Token embeddings: Q2_K
• Logical code bits/weight: 2.0000
• Physical container bits/weight: 2.5625
• Serialized size: 2,635,548,156 bytes including manifest.json
• Tensor histogram: IQ2_S=7,568,097,280, Q2_K=622,329,856, F32=308,224

Integrity

• GGUF SHA-256: 72da28b975863b66b69247fb7b1d74a925b5ea4d841a6034f3c4a115f9137235
• Manifest SHA-256: 6beff228c81a55d07051128689c297bf5e87867a5c6cead9330d3c8c8009f64b
• Ralph content hash: 7de45062085db016bb90f7bddd204367b19683bb99ca893dad009f0a42dc7230

The Ralph v2 inspect_checkpoint, sub2 bit-tier, and pinned-parent compatibility gates all pass on the exact packaged directory.

Runtime checks

• llama-cpp-python==0.3.34, CUDA backend, validator settings (n_ctx=8192, n_batch=512, all layers offloaded): loaded and generated successfully.
• Two independent generations over the same 72 signed round-2 prefixes were identical for all 72 outputs.
• Ralph degeneracy gate: pass, 21/72 outputs above the six-gram repetition threshold (29.17%; formal rejection is above 30%).
• Native llama.cpp e3546c7948e3af463d0b401e6421d5a4c2faf565 bounded forward pass: exit 0, one-chunk calibration perplexity 1.4308.

Observer replay

Diagnostic miner-side replay against the 72 post-commit indices, frozen parent steps, frozen continuations, and Phi-3 observer from signed public round 2:

Worst-slice retention was 0.17885, with 72/72 samples scored, zero discards, mean effect 0.6402, and no scorer reasons. This is an unsigned miner diagnostic, not a validator verdict or crown certificate.