---
base_model: nbeerbower/Qwen3.6-27B-TIES
library_name: peft
license: apache-2.0
datasets:
- schneewolflabs/egirl-delegation-dpo
language:
- en
tags:
- lora
- orpo
- peft
- qwen3.6
- agents
- tool-use
---

# Qwen3.6-27B-delegation-LoRA

An ORPO LoRA that teaches a local operator model **when to hand work to a coding agent instead of
poking at the codebase itself**.

Trained on [`schneewolflabs/egirl-delegation-dpo`](https://huggingface.co/datasets/schneewolflabs/egirl-delegation-dpo)
for [egirl](https://github.com/nbeerbower/egirl), where the local model is meant to be the
human-in-the-loop and delegate engineering work to `code_agent` (Claude Code / Codex).

**Delegation goes from 0/3 to 3/3.**

## Results

Ten operator requests in egirl's `<tool_call>` dialect, with its real system prompt:

| metric | `Qwen3.6-27B-TIES` | **+ this LoRA** |
|---|---|---|
| tool + args correct | 7/10 | **8/10** |
| **delegated coding work** | **0/3** | **3/3** |
| valid tool calls | 10/10 | 10/10 |
| no hallucinated tools | 10/10 | 10/10 |
| terminated | 10/10 | 10/10 |

All three delegation cases flipped from exploring to handing off:

| request | before | after |
|---|---|---|
| "split the auth module, make sure tests pass" | `glob_files`, `glob_files` | **`code_agent`** |
| "fix the Discord 2000-char bug" | `glob_files`, `glob_files` | **`code_agent`** |
| "add retry logic to every HTTP call, plus tests" | `glob_files`, `glob_files` | **`code_agent`** |

### Why this needed training

Delegation is not something prompting fixes. Every model benched scored **0/3**, including with a
system prompt that says *"Delegate coding work to the code agent by default — you're the
human-in-the-loop, not a code generator"*:

| model | delegated |
|---|---|
| `Qwen/Qwen3.6-27B` | 0/3 |
| `Qwen3.6-27B-TIES` | 0/3 |
| `Huihui-ThinkingCap-Qwen3.6-27B-abliterated` | 0/3 |
| `nightmedia/…-Architect-Polaris2-Fable-B-F451` | 0/3 |
| two other local merges | 0/3 |
| **`gpt-5.6-sol`** | **1/3** |

Swapping a generic bench prompt for egirl's real one moved overall tool accuracy 6/10 → 7/10 and
changed delegation not at all.

### Regressions

Two cases got *more* eager to act, which is the expected failure direction for this axis:

| case | before | after |
|---|---|---|
| "what's in /etc/hostname?" | `read_file` | `execute_command` |
| "in one sentence, TIES or SLERP?" | no tool (correct) | `web_research` |

The dataset has a 45-row band of single-line-edit examples meant to hold the "just do it yourself"
boundary, against 121 delegate rows. That ratio looks too delegate-heavy; a rebalanced mix is the
obvious next iteration.

**Censorship eval also shifted** on the same base — `ccp_truth_neutral` 4.00/4 → 2.40/4, total
29.00 → 26.60. Training on tool-use is not supposed to touch that axis, so treat it as evidence
that a narrow LoRA at r=32 still perturbs unrelated behaviour. Stack with
[`Bubba-3ep`](https://huggingface.co/nbeerbower/Bubba-Qwen3.6-27B-LoRA-3ep) if you want the
censorship behaviour back, and re-measure.

## Usage

llama.cpp, no merging required:

```bash
python convert_lora_to_gguf.py --base /path/to/Qwen3.6-27B --outtype f16 \
    --outfile delegation-lora-f16.gguf /path/to/Qwen3.6-27B-delegation-LoRA

llama-server -m Qwen3.6-27B-TIES-Q8_0.gguf --lora delegation-lora-f16.gguf -ngl 99 --jinja
```

PEFT:

```python
from peft import PeftModel
from transformers import AutoModelForImageTextToText

model = AutoModelForImageTextToText.from_pretrained("nbeerbower/Qwen3.6-27B-TIES", dtype="bfloat16")
model = PeftModel.from_pretrained(model, "nbeerbower/Qwen3.6-27B-delegation-LoRA")
```

## Training

| | |
|---|---|
| method | ORPO (β=0.1) |
| base | `nbeerbower/Qwen3.6-27B-TIES` |
| data | 558 pairs (186 tasks × ~3 workspace-path variants) |
| LoRA | r=32, α=64, dropout 0.0 |
| targets | `q,k,v,o,gate,up,down_proj` |
| lr | 8e-6 cosine, 10% warmup |
| batch | 2 × 4 accum (effective 8) |
| epochs / steps | 2 / 138 |
| loss | 1.970 → 0.659 |
| hardware | DGX Spark (GB10), bf16, gradient checkpointing |

Trained with [Merlina](https://github.com/Schneewolf-Labs/Merlina).

The training prompt is deliberately **minimal** — identity, workspace, tool definitions, nothing
else. No delegation policy and no "call the tool immediately" instruction, so the model learns
*when* to delegate as behaviour rather than learning to obey a sentence in context. egirl's real
prompt does state the policy, and the two compose.

## Limitations

- **Bound to egirl's tool schema.** The `chosen` targets name `code_agent`, `glob_files`,
  `git_status` and friends. A different tool set needs different data.
- **Single-turn.** It teaches the first action only — not what to do with results, when to follow
  up, or how to recover from a failed call.
- **186 underlying tasks**, each seen under ~3 workspace paths. Path memorisation is addressed;
  task diversity is not.
- **Over-eagerness**, and the censorship shift noted above.

## Related

- [`schneewolflabs/egirl-delegation-dpo`](https://huggingface.co/datasets/schneewolflabs/egirl-delegation-dpo) — the training data
- [`schneewolflabs/egirl-hemlock-dpo`](https://huggingface.co/datasets/schneewolflabs/egirl-hemlock-dpo) — companion axis, writing Hemlock through a tool call
- [egirl](https://github.com/nbeerbower/egirl) — the agent this targets