ganesh-review

ganesh-review

An experimental Qwen3.5-2B adapter for predicting software-change consequences from structured state/action requests. Author: Navneet Prabhakar.

This is a Qwen-derived adapter, not a new foundation model, not the independent Laxmi world-model deliverable, and not a production code reviewer. The name does not imply that raw repositories or pull requests are supported inputs.

Release status

Public Hugging Face repository: adminspec/ganesh-review. The first verified public artifact revision is 9c4b5c1b9412d176424a66ba2b7667f079f21185. See release status.

What is preserved

• Original custom rank-8, alpha-16 adapter: 7,815,168 trained parameters across 150 modules; 300 tensors; 31,293,440 bytes.
• Original model/inference configuration and training metadata.
• Extracted training implementation, including checkpoint/resume support.
• A portable inference entry point under qualification, artifact checks, and a technical report documenting successes and failures.

The historical format ID still contains LAXMI; retaining it preserves artifact identity. Renaming the project does not alter the weights or claim new training.

Results and limitations

Training used 432 records, two epochs and one seed. There is no demonstrated general code-review reliability, broad generalization, multi-seed robustness, or novel world-model mechanism. The private cases/oracles are not distributed, so the exact private result cannot be reproduced from this repository alone. See technical report.

Inspect without a model download

python verify_artifacts.py
python -m unittest discover -s tests -v

Tensor tests require PyTorch and safetensors. The verifier uses only the standard library. No command above trains a model or calls a cloud service.

Merged-model inference

The Hugging Face repository root contains a merged BF16 safetensors model. It was reloaded with AutoModelForCausalLM and completed a finite-logit forward smoke on CPU. This validates file loading and execution, not model quality.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "adminspec/ganesh-review"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16)

Apply the prompt framing from adapters/custom/inference-config.json; the model expects a canonical structured S4 request, not a raw repository or pull request.

Adapter formats

The original custom adapter and a mechanically converted PEFT adapter are under adapters/. The conversion loaded all 300 tensors into PEFT, saved them, and verified exact equality after a fresh adapter reload. The current upstream Qwen weight file has a different hash from the exact training base; therefore the merged root model is the strongest self-contained consumption path. Loading the PEFT adapter against a newer upstream base is compatibility use, not a claim of evaluated numerical identity.

With a separately acquired, verified local base snapshot and an admissible canonical S4 JSON object:

python src/infer.py --model /absolute/path/to/base-snapshot --input request.json

This local command requires CUDA, uses local files only, verifies the historical base weight hash, and emits unmodified generated text. It does not repair invalid JSON or claim semantic validation.

GGUF

No GGUF is published. The latest official llama.cpp conversion scripts checked during release preparation do not support the qwen3_5 architecture. Renaming another Qwen architecture or forcing an unsupported conversion would not be a consumable artifact. GGUF can be added later after upstream converter support and an actual load/generation validation.

Training

training/train_lora.py preserves the source used by the retained execution snapshot. It is a historical Linux/CUDA trainer with explicit UID/GID 65532, root-owned input and offline-environment checks, not a plug-and-play laptop training command. Its --help is safe to inspect. The training corpus and its generator dependency closure have not been released; exact retraining is not currently reproducible from this repository alone. No additional training is part of the release preparation.

Rights and provenance

The base is published by Qwen under Apache-2.0. That does not assign a license to these contributions or clear the training data. Owner license approval and source/data provenance review remain explicit release gates. No employer affiliation or endorsement is claimed.

No Laxmi Git history, private evaluation data, recovery keys, cloud credentials, operational backups or optimizer checkpoints are included.