EdgeGuard Cypher Qwen3 4B v0.8 Feedback-Validated GGUF
Private GGUF export for EdgeGuard direct text-to-Cypher edge-node runtime use.
This preview continues EGM-026 v0.8 feedback-validated LoRA.
Artifact
- File:
edgeguard-cypher-qwen3-4b-v0.8-feedback-validated.Q4_K_M.gguf
- Format: GGUF
- Quantization:
Q4_K_M
- SHA256:
1da636448aafe9f2ff1826008e5f6f10ccd33488172c386f84783b0f5b485f77
- Size bytes:
2497278816
Source
- Base model:
Qwen/Qwen3-4B-Instruct-2507
- Base revision:
cdbee75f17c01a7cc42f958dc650907174af0554
- Source adapter SHA256:
9dced541868ccb6a8977591b8e79bc1ce4824f2cacebb0e369366ac529addf8d
- Training target: assistant output is one valid read-only Cypher query string only.
Method
Merged EGM-026 QLoRA adapter into pinned Qwen/Qwen3-4B-Instruct-2507 revision cdbee75f17c01a7cc42f958dc650907174af0554, converted to BF16 GGUF, then quantized to Q4_K_M for edge-node.
Evaluation Summary
Best raw EdgeGuard Cypher model so far on the frozen v0.8 bundle: validation subgraph accepted 99/100, sealed-test subgraph accepted 98/100, probe subgraph accepted 197/200, scalar regressions 0. Generated-live hard success is 173/180 = 96.1% when no-graph rows are non-blocking; no-graph rows are 13.
Promotion Status
Private edge-node GGUF publication. No-graph live results are accepted as empty/coverage results; hard live failures remain 4 planner failures and 3 execution failures. Use behind the EdgeGuard schema/read-only validator.
Runtime
The intended consumer is edge-node LLM_INFERENCE_API through
LlamaCppBaseServingProcess, using either Hugging Face MODEL_NAME +
MODEL_FILENAME or a mounted local MODEL_PATH.
Recommended edge-node values:
1AI_ENGINE=edgeguard_qwen_4b
2MODEL_NAME=ratio1/edgeguard-cypher-qwen3-4b-v0.8-feedback-validated-gguf
3MODEL_FILENAME=edgeguard-cypher-qwen3-4b-v0.8-feedback-validated.Q4_K_M.gguf
This model must be used behind the EdgeGuard schema/read-only validator. Do not
execute raw model output directly against Neo4j.
