Verified Patches for Qwen/Qwen3.5-0.8B

Browse pre-registered, Ed25519-signed behavioral and configuration patches for Qwen/Qwen3.5-0.8B. Fix thinking loops, generation errors, and repetition penalty issues without downloading 16 GB forks.

Available Patches (2)

P1_WEIGHTS wsp-qwen3.5-0.8b-math-loop-exp8

Qwen3.5-0.8B 32k Thinking-Loop Surgical Cure (MIT CSAIL Exp-8)

Eliminates infinite 32,768-token chain-of-thought looping in Qwen/Qwen3.5-0.8B by modifying 64 output columns of mlp.down_proj in layers 15 and 19 (131,072 bf16 weights, 0.016% of parameters). Lifts independent MATH-500 accuracy from 5.0% (7.9% paper) to 51.0% (48.1% paper).

$ pip install wsphub && wsp apply Qwen/Qwen3.5-0.8B --patch wsp-qwen3.5-0.8b-math-loop-exp8 --out ./patched
P0_CONFIG wsp-qwen3.5-0.8b-missing-genconfig-eos-fix

Qwen3.5-0.8B Missing generation_config & Dual Stop-Token Fix

Adds the missing generation_config.json to Qwen/Qwen3.5-0.8B with dual EOS stop tokens [248046 (<|im_end|>), 248044 (<|endoftext|>)] and verified sampling parameters, curing post-turn hallucination bleed.

$ pip install wsphub && wsp apply Qwen/Qwen3.5-0.8B --patch wsp-qwen3.5-0.8b-missing-genconfig-eos-fix --out ./patched

Verified Patch Catalog (P0 Config & P1 Weights)

Every patch is pre-registered, evaluated across 10,000 paired bootstrap and permutation tests, checked against the secret perturbed WSP-Holdout split and StrongREJECT/XSTest safety gates, and signed with Ed25519.

P1_WEIGHTS wsp-qwen3.5-0.8b-math-loop-exp8 Target: Qwen/Qwen3.5-0.8B

Qwen3.5-0.8B 32k Thinking-Loop Surgical Cure (MIT CSAIL Exp-8)

Cures the catastrophic 93.6% 32k-token repetition loop on MATH-500 through 64 targeted delta columns across layers 15 & 19 (+48.1% MATH-500 accuracy uplift).

$ pip install wsphub && wsp apply Qwen/Qwen3.5-0.8B --patch wsp-qwen3.5-0.8b-math-loop-exp8 --out ./patched
P0_CONFIG wsp-qwen3.5-0.8b-missing-genconfig-eos-fix Target: Qwen/Qwen3.5-0.8B

Qwen3.5-0.8B Missing generation_config & Dual Stop-Token Fix

Restores canonical stop tokens and anti-loop presence penalties in generation_config.json.

$ pip install wsphub && wsp apply Qwen/Qwen3.5-0.8B --patch wsp-qwen3.5-0.8b-missing-genconfig-eos-fix --out ./patched
P0_CONFIG wsp-qwen2.5-coder-1.5b-rep-penalty-syntax-fix Target: Qwen/Qwen2.5-Coder-1.5B-Instruct

Qwen2.5-Coder-1.5B Repetition Penalty Syntax Truncation Neutralization

Neutralizes repetition penalty=1.1 identifier/brace corruption in code & JSON generation.

$ pip install wsphub && wsp apply Qwen/Qwen2.5-Coder-1.5B-Instruct --patch wsp-qwen2.5-coder-1.5b-rep-penalty-syntax-fix --out ./patched

WSP Verify CI/CD & Supernova Settlement Rail

Automated regression testing and attestation for model builders, enterprise publishers, and AI coding agents:

Streamable HTTP MCP Server Endpoint

Direct JSON-RPC 2.0 integration for Cursor, Windsurf, Claude Code, and autonomous AI agents:

Endpoint: POST https://api.wsphub.dev/api/v1/mcp
Discovery Tool: wsp.search_verified_patches (Free)
Metered Gateway: https://supernova.cool/api/v1/mcp/ (X-Supernova-Key)

Frequently Asked Questions (FAQ)

What is WSP (Weight Sparse Patch) format?

WSP is an open binary specification (.wsp v2.1) that stores surgical weight deltas and configuration fixes for open-weight models. It allows instant application in under 5 seconds via Copy-on-Write Safetensors without redownloading 16 GB base models.

How does WSP achieve 94% to 99% smaller files than LoRA?

LoRA applies low-rank matrices across all layers. WSP isolates the exact causal columns responsible for bugs (e.g. 64 columns across 2 MLP layers), reducing the file size down to 0.8 MB – 4 MB.

Can patches be reverted bit-exactly?

Yes. Running wsp revert <dir> restores the pristine original SHA-256 weight state in <1 second.