README.md

---
base_model: Qwen/Qwen2.5-7B-Instruct
library_name: peft
license: mit
tags:
  - adele
  - demand-level-annotation
  - lora
  - qlora
  - distillation
---

# Distilled Qwen2.5-7B ADeLe Demand-Level Annotator (QLoRA adapter)

A QLoRA-fine-tuned adapter for `Qwen/Qwen2.5-7B-Instruct` that predicts
**ADeLe v1.0 / DeLeAn demand-level rubric scores** (18 rubrics, 0-5 scale),
distilled from three large annotator LLMs (Qwen2.5-72B-Instruct-AWQ,
Llama-3.3-70B-Instruct-AWQ, Mistral-Large-Instruct-2411-AWQ).

Full project, training/eval scripts, and results:
https://github.com/ceferra/adele-annotation-distillation

Based on the official ADeLe project:
https://github.com/Kinds-of-Intelligence-CFI/ADeLe-AIEvaluation

## Training data

Official ADeLe v1.0 battery (16,108 items, 63 tasks). Tasks were split
51 (train) / 6 (validation) / 6 (test), by task name (seed 42), so that
validation/test tasks are never seen during training — see
`docs/split_manifest.json` in the main repo for the exact task lists.

## Training hyperparameters

- Base model: `Qwen/Qwen2.5-7B-Instruct`
- Method: QLoRA, 4-bit NF4 quantization
- LoRA: r=16, alpha=32, dropout=0.05
- Target modules: `q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj`
- Max sequence length: 4096
- Epochs: 1.0
- Learning rate: 2e-4
- Effective batch size: 4 GPUs x 1 (per-device) x 16 (gradient accumulation)

## Evaluation summary

Against the actual GPT-4o annotations present in the project's held-out
benchmark data (5 external benchmarks: MBPP, TheoremQA, Omega, RelBench,
SWE-bench-Verified; rWG within-group agreement index, mean over 18
rubrics):

| Annotator | rWG vs. GPT-4o |
|---|---|
| Qwen2.5-72B-Instruct-AWQ | 0.892 |
| Llama-3.3-70B-Instruct-AWQ | 0.908 |
| Mistral-Large-Instruct-2411-AWQ | 0.918 |
| **Distilled Qwen2.5-7B (this adapter)** | **0.934** |
| Median ensemble of all 4 | 0.944 |

The distilled 7B model matches or exceeds each individual large model's
agreement with GPT-4o, at a small fraction of the inference cost, on 17 of
18 rubrics (the exception being rubric KNa, where all candidates, including
the full ensemble, score below 0.78).

See the main repository's `results/rwg_gpt4o_analysis/` and
`reports/DeLeAn_Full_Project_Report.pdf` for full details and all 18
per-rubric numbers.

## How to use

```python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-7B-Instruct")
model = PeftModel.from_pretrained(base, "path/to/this/adapter")
tok = AutoTokenizer.from_pretrained("path/to/this/adapter")
```

For fast batch inference, merge the adapter into the base model first
(see `scripts/merge_lora.py` in the main repository) and serve with vLLM.

## License

MIT. Please cite the ADeLe paper (Zhou et al., Nature 2026;
arXiv:2503.06378) and the official ADeLe repository
(https://github.com/Kinds-of-Intelligence-CFI/ADeLe-AIEvaluation) if you use
this adapter, since it is trained entirely on ADeLe's official task battery.

### Framework versions

- PEFT 0.13.2