FlowSpec Qwen3-1.7B model and experiment artifacts.
Included assets
flowspec-qwen3-1.7b-qlora-adapter.tar.gz: final LoRA adapter and tokenizer files.flowspec-qwen3-1.7b-q4_k_m.gguf: merged GGUF Q4_K_M model for llama.cpp CPU inference.flowspec-experiment-evidence.tar.gz: raw predictions plus selected training, export, and end-to-end logs.SHA256SUMS: checksums for all three archives/model files.
Verified result snapshot
The final QLoRA run reached 1.000 Schema validity, 1.000 DAG validity, and 1.000 sandbox pass rate on the fixed 250-example test set. Semantic structure score was 0.7930, +0.1195 over few-shot + validator. A 5-node Chinese workflow also passed the GGUF llama.cpp -> FastAPI -> validator -> sandbox path.
The datasets are original synthetic workflows. Metrics remain provisional until the documented author review is complete. See MODEL_CARD.md, reports/experiment-summary.md, and docs/error-analysis.md in the repository.
Base model: Qwen/Qwen3-1.7B. License: Apache-2.0.