metadata
license: apache-2.0
base_model: Qwen/Qwen3.5-35B-A3B-Base
tags:
- lora
- fine-tuned
- tool-calling
- mcp
- dbt
ecu-pilot (FP16)
Fine-tuned Qwen3.5-35B-A3B-Base for structured tool calling against project metadata via MCP.
Trained to accurately call 9 tools — lineage traversal, impact analysis, test coverage reporting, schema introspection, search, and more — with valid arguments and well-synthesized answers grounded in real tool output.
Model details
| Base model | Qwen3.5-35B-A3B-Base |
| Architecture | Mixture of Experts (35B total, 3B active per token) |
| Fine-tuning method | bf16 LoRA (r=16, alpha=16) |
| Training stages | Stage 1: tool mechanics (1 epoch, 1,206 examples) / Stage 2: structured planning (2 epochs, 290 examples) |
| Hardware | NVIDIA H200 141GB, ~1 hour total |
| Training data | 1,206 ChatML examples with real tool responses from indexed project metadata |
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model = AutoModelForCausalLM.from_pretrained(
"mach-kernel/ecu-pilot-fp16",
torch_dtype=torch.bfloat16,
device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("mach-kernel/ecu-pilot-fp16")
Quantized variants
| Format | Repository |
|---|---|
| FP16 (this repo) | mach-kernel/ecu-pilot-fp16 |
| LoRA adapter only | mach-kernel/ecu-pilot-fp16-lora |
| GGUF Q4_K_M | mach-kernel/ecu-pilot-q4km |
| GGUF Q8_0 | mach-kernel/ecu-pilot-q8_0 |
Training methodology
Two-stage supervised fine-tuning adapted from the Thinkquel methodology:
- Stage 1 — Tool mechanics: Teaches the model what tools exist, how to format calls, and how to interpret responses.
- Stage 2 — Structured planning: Teaches the model to reason about when and why to call tools using
<think>blocks before acting.
All training examples use real tool responses from an indexed project — no synthetic or hallucinated tool output.
Why "ecu"
No particular reason. Just liked the sound of it.
Why ecu
No reason. Just liked how it sounded. Definitely not a Caesar cipher of anything. Don't look into it.