SLM Workflow Planner v3 — PEFT Format (Universal)
HuggingFace PEFT format — works on any platform (CUDA, CPU, Apple Silicon).
Converted from MLX LoRA adapter (MLX version).
Quick Start
Model Details
| Property | Value |
|---|---|
| Base Model | Qwen/Qwen2.5-7B-Instruct |
| LoRA Rank | 16 |
| LoRA Alpha | 32 |
| Target Modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| Layers | 0-27 (28 layers) |
| Task | Workflow execution planning (NEXT/RETRY/FORK/JOIN/META) |
Performance (Solo)
| Category | Score |
|---|---|
| NEXT | 12/22 (55%) |
| RETRY | 12/12 (100%) |
| FORK | 14/14 (100%) |
| JOIN | 15/15 (100%) |
| META | 0/13 (0%) |
| Total | 53/76 (69.7%) |
Ensemble Performance (v3 + v7 + GPT-4.1 arbiter)
| Category | Score |
|---|---|
| NEXT | 19/22 (86%) |
| RETRY | 11/12 (92%) |
| FORK | 14/14 (100%) |
| JOIN | 15/15 (100%) |
| META | 8/13 (62%) |
| Total | 67/76 (88.2%) |
Role in Ensemble
v3 is the structural expert — perfect on RETRY, FORK, and JOIN decisions. In the 3-expert ensemble, it handles all structural topology decisions while v7 handles META/boundary cases and GPT arbitrates disagreements.
- Downloads last month
- 35
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support