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

Status: candidate for Hugging Face upload, pending operator authorisation.

This adapter is the first measured Sounio fine-tuning artefact in this lane that
beats the zero-training base-model reference on the full 200-row held-out eval
split.

## Identity

| Field | Value |
|---|---|
| Base model | `Qwen/Qwen2.5-Coder-1.5B` |
| Training objective | Sounio complete-file SFT with surgical repair pairs |
| Branch | `feature/dataset-expansion` |
| Dataset commit | `12c471d1a` |
| Run ID | `sounio-qwen15-lora-full200-surgical-repair-20260522T021535` |
| Slurm job | `1603` |
| GPU | 1 GPU on `gpu-orangefs` |
| Training rows requested | 4,800 |
| Training rows used | 4,769 |
| Optimiser steps | 120 |
| Prompt style | `sounio-compact-completion` |
| Repair pair ratio | 0.20 |
| Run-expected ratio | 0.10 |
| Separate runtime repair ratio | 0 |
| Targeted runtime ratio | 0 |

## Artifact Location

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

Adapter files present:

| File | Size |
|---|---:|
| `adapter_model.safetensors` | 73,911,112 bytes |
| `adapter_config.json` | 797 bytes |
| `README.md` | 5,097 bytes |
| `tokenizer.json` | 11,421,896 bytes |
| `tokenizer_config.json` | 7,338 bytes |
| `vocab.json` | 2,776,833 bytes |
| `merges.txt` | 1,671,853 bytes |
| `special_tokens_map.json` | 616 bytes |
| `added_tokens.json` | 605 bytes |

Checksums:

```text
9d8540e218439bb28026115fbe8d4be5fe18162a0a1c07474d9850579ee36c4a  adapter_model.safetensors
0ce16cbdfa4be1b4fb97868ee0b5f2fed33dfb3ba65c7961530f7fe7cd544057  train_summary.json
6eca4fdaaf6e3bb48f302f8090e681cddb3fc6e08b7a99dcde0f04962c09d794  lora_smoke_eval_report.json
```

## Evaluation Result

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%) |
| Qwen2.5-Coder-1.5B LoRA candidate | 192/200 (96.0%) | 192/200 (96.0%) | 58/65 (89.2%) |
| Candidate + structural repair layer | 199/200 (99.5%) | 198/200 (99.0%) | 59/65 (90.8%) |

The adapter improves the two acceptance metrics used for this lane:

- `souc check` pass rate on the full held-out eval split
- exact stdout match on all runnable held-out rows

This is a utility result, not just a training-loss result. The generated outputs
extract as complete Sounio files, mostly satisfy the no-prose/no-Rust contract,
compile with `souc check`, and run correctly on most rows with expected stdout.

The 98%+ publication-confidence run is adapter-only: it reuses the saved adapter
and widens the deterministic structural repair layer; it does not retrain model
weights. Result path:

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

The structural repair layer catches outputs still outside the complete-file
objective, including Rust-shaped `String`/`Vec` surfaces, invalid synthetic
declarations, algebra blocks used as structs, and one observed scalar typo.

## Remaining Misses

Residual failures after the 98%+ adapter-only evaluation:

- compile or contract failures: 2/200
- runtime misses: 6/65 runnable rows
- residual Rust-shaped surface:
  - `String`: 1 row
- runtime families still missing:
  - one HumanEval runnable
  - two FFI logical extent rows printing `16`
  - two GPU tile rows printing `4`
  - one reserve advanced row printing `64`

These are the next repair-pair targets if the goal is to improve this adapter
before publication.

## Upload Policy

Do not upload automatically. Upload requires explicit operator authorisation and
an operator-owned Hugging Face token.

Suggested repository name:

```text
chiuratto-AIgourakis/sounio-qwen25-coder-15b-lora
```

Suggested upload payload:

```text
adapter/
train_summary.json
lora_smoke_eval_report.json
lora_smoke_predictions.jsonl
ADAPTER_CANDIDATE_QWEN15.md
```

The adapter should be described as an experimental Sounio LoRA adapter, not as a
general-purpose production code model.