---
license: apache-2.0
base_model: Qwen/Qwen2.5-Coder-1.5B
library_name: peft
tags:
  - sounio
  - code
  - peft
  - lora
  - qwen2.5-coder
language:
  - en
pipeline_tag: text-generation
---

# Sounio Qwen2.5-Coder-1.5B LoRA Adapter

Experimental LoRA adapter for generating complete Sounio source files.

This is not a general-purpose production code model. It is a research adapter
trained for the Sounio dataset-expansion lane, where the target is complete,
compiler-checkable `.sio` programs rather than prose continuations.

## Base Model

`Qwen/Qwen2.5-Coder-1.5B`

## Training Snapshot

| Field | Value |
|---|---:|
| Dataset branch commit | `aa5f9218e` |
| Training rows requested | 4,800 |
| Training rows used | 4,769 |
| Optimiser steps | 120 |
| Repair-pair rows | 480 |
| Run-expected rows | 435 |
| Prompt style | `sounio-compact-completion` |
| LoRA rank | 16 |
| LoRA alpha | 32 |

## Evaluation

Held-out split: `datasets/hf_examples/instruction_pairs_eval.jsonl`, 200 rows.

| Model/objective | Compile pass | Compile + contract-clean | Exact runtime pass |
|---|---:|---:|---:|
| Base model, zero training | 111/200 (55.5%) | not recorded | 25/65 (38.5%) |
| Adapter candidate | 192/200 (96.0%) | 192/200 (96.0%) | 58/65 (89.2%) |
| Adapter + structural repair layer | 199/200 (99.5%) | 198/200 (99.0%) | 59/65 (90.8%) |

The 98%+ result uses the saved adapter plus the deterministic structural repair
layer in the Sounio repository. The repair layer catches outputs that are still
outside the complete-file objective, such as Rust-shaped `String`/`Vec`
surfaces, invalid synthetic declarations, algebra blocks used as structs, and
one observed scalar typo. It does not retrain model weights.

## Artefacts Included

- `adapter_model.safetensors`
- `adapter_config.json`
- tokenizer files copied with the adapter
- `train_summary.json`
- `lora_smoke_eval_report.json`
- `adapter_eval_summary.json`
- `adapter_eval_report.json`
- `ADAPTER_CANDIDATE_QWEN15.md`

## Reproducibility Notes

Primary adapter training run:

```text
/orangefs/training/sounio/hf-examples/lora-smoke/sounio-qwen15-lora-full200-surgical-repair-20260522T021535/results
```

98%+ adapter-only evaluation run:

```text
/orangefs/training/sounio/hf-examples/lora-smoke/sounio-qwen15-adapter-eval-98plus-20260522T083913/results
```

Key checksum:

```text
9d8540e218439bb28026115fbe8d4be5fe18162a0a1c07474d9850579ee36c4a  adapter_model.safetensors
```

## Intended Use

Use this adapter to generate or repair small Sounio examples under the
`sounio-compact-completion` prompt format. Outputs should still be validated
with `souc check` and, when applicable, runtime checks.