moinonin/qwen2.5-coder-7b-promex

Model Overview

moinonin/qwen2.5-coder-7b-promex is a fine-tuned version of the Qwen2.5-Coder-7B-Instruct model, specifically adapted using LoRA adapters on the Spec-Forge training corpus. Its primary function is to transform natural language feature requests into structured, validated YAML specifications that adhere to the COMMAND_RUNWAY methodology.

Key Capabilities

• YAML Specification Generation: Converts user prompts into detailed YAML specifications, including task_id, summary, depends_on, local_goals, global_goals_refs, and context.
• Structured Verification Flows: Each local_goal within the generated spec includes an Inspect → Create/Modify → Verify verification flow.
• Hardened Validation: Specs are designed to pass a robust validator, ensuring canonical vocabulary, near-duplicate detection, and YAML safety.
• Runbook Readiness Scoring: Generated specifications are scored against runbook-readiness criteria, with a hard gate for missing verification stages.

Training Details

The model was fine-tuned for 3 epochs using a LoRA rank of 16 and an effective batch size of 4. It leverages 4-bit NF4 quantization and was trained on a synthetic corpus of 475 seed prompts across 21 feature categories. The training data was generated by Ollama (qwen2.5-coder:7b-instruct) and rigorously validated against a hardened YAML spec validator and a runbook scorer.

Limitations

• Quality is bounded by the base model's spec-generation ability, as it was trained on a synthetic corpus generated by the base model itself.
• Specs are scoped to single-file, single-feature granularity, not multi-stage epics.
• Context is fixed to a TypeScript/Express/Prisma/Vitest stack.
• GGUF quantization (q4_k_m) may introduce minor quality degradation compared to the 16-bit merged model.