# Release Audit

Generated: 2026-06-28T00:58:55.373853+00:00

- HF repo: `LordNeel/Ornith-1.0-35B-GGUF-llamacpp-tp1`
- Remote head before this upload: `16abb00acc4a1c89dbc5c357252fe677205931cd`
- Remote files before this upload: 241
- Staged support files excluding generated audit files: 217
- Overall status: `pass for serving/quant/MTP support files; SFT adapter remains unreleased`
- Latest HF release verifier: `pass` (`25` checks, `0` failures, `1` expected SFT warning)
- Latest serving profile validation: `pass` (`43/43` checks)
- Latest behavior regression matrix: `5` passing artifacts, `2` known-bad failing artifacts
- Latest release readiness: serving release `pass`, full project `incomplete`

## Model Card Metadata

| Field | Value |
|---|---|
| `license` | `mit` |
| `license_link` | `https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B/blob/main/LICENSE` |
| `base_model` | `deepreinforce-ai/Ornith-1.0-35B` |
| `pipeline_tag` | `text-generation` |
| `library_name` | `gguf` |
| `tags` | `gguf, llama-cpp, quantized, text-generation, coding, mtp, qwen3_5_moe, single-gpu, tp1` |

## Current Recommendations

- Speed-focused serving: `Q4_K_M` through llama.cpp with `REASONING=off`.
- Lowest-memory validated GGUF: `Q3_K_M`.
- Middle-footprint validated GGUF: `IQ4_XS`.
- Integrated MTP artifact: `ornith-1.0-35b-IQ4_XS-MTP-graft-headQ6.gguf`.
- Future SFT input if explicitly resumed: `data/train/coding_sft_prod_repair_strong.jsonl` regenerated locally from `configs/sft_repair_blend.yaml`.
- Required future resume guard: `scripts/preflight_sft_resume.py`.
- Multi-GPU serving remains out of scope; all serving profiles are single-GPU `tp=1`.

## SFT Remediation Update

| Item | Result | Evidence |
|---|---:|---|
| Code-only extracted train rows | 20000 | `probes/sft-code-only-extract.json` |
| Recommended strong train rows | 44096 | `probes/sft-data-validation-prod-repair-strong.json` |
| Strong validation blocking issues | 0 | `probes/sft-data-validation-prod-repair-strong.json` |
| Strong train assistant code fences | 54.4267% | `probes/sft-format-audit-prod-repair-strong.json` |
| Strong train assistant `Here's` rate | 30.4948% | `probes/sft-format-audit-prod-repair-strong.json` |
| Code-only format gate | 8/8 pass | `probes/sft-format-gate-code-only-extract.json` |
| Strong repair format gate | 10/10 pass | `probes/sft-format-gate-prod-repair-strong.json` |
| Resume preflight | 16/16 pass | `probes/sft-resume-preflight-strong.json` |

## SFT Status

- The production LoRA SFT run was stopped at user request.
- Latest durable checkpoint: `checkpoint-8000`.
- No SFT adapter is released from this run.
- The resume preflight is report-only and does not launch training.

## HF Release Verification

- Script: `scripts/verify_hf_release.py`
- Latest report: `probes/hf-release-verification.json`
- Verified remote commit: `7dcf02d87d07e4b173d7b375e9fc747605d80652`
- Result: `25` checks passed, `0` failures, `1` expected SFT warning.
- Coverage: model-card metadata, required support files, staged support-file
  hashes, all `14` published GGUF artifacts, SFT remediation gate summaries,
  behavior-matrix summary, and release-readiness summary.

## Behavior Regression Matrix

- Script: `scripts/summarize_behavior_matrix.py`
- Latest report: `probes/behavior-regression-matrix.json`
- Markdown summary: `probes/behavior-regression-matrix.md`
- Current run: `7` artifacts, `98` checks, `5` passing artifacts, `2` failing
  known-bad artifacts, `9` failed checks, and `6` blocking issues.
- Known-bad artifacts are intentionally included: the old reasoning-on Q4
  profile and the stopped checkpoint-8000 LoRA adapter.

## Serving Profiles

- Catalog: `configs/serving_profiles.yaml`
- Validator: `scripts/validate_serving_profiles.py`
- Latest report: `probes/serving-profile-validation.json`
- Markdown summary: `probes/serving-profile-validation.md`
- Current run: `43/43` checks passed, `0` failures, `0` warnings.
- Profiles: Q4_K_M speed/default, Q3_K_M lowest-memory, IQ4_XS
  middle-footprint, and adaptive IQ4_XS MTP graft.

## Release Readiness

- Script: `scripts/build_release_readiness_report.py`
- Latest report: `probes/release-readiness-report.json`
- Markdown summary: `probes/release-readiness-report.md`
- Current run: `25` checks, `0` failures, `5` warnings.
- Serving GGUF release status: `pass`.
- Full project status: `incomplete` because no SFT/post-training adapter is
  released.

## Machine-Readable Manifest

See `ARTIFACT_MANIFEST.json`.