depthfirst 路 Nemotron 3.5 distilled from DFS Large
An experimental rank-8 LoRA checkpoint of NVIDIA Nemotron 3.5 Lightning, trained for 20 optimizer updates using depthfirst鈥檚 DFS Large teacher. A selected five-task pilot is complete. Broader evaluation is still in progress; this release does not establish a general improvement over the base model.
Before-and-after traces 路 Interactive comparison 路 Training metrics
Checkpoint
| Property | Value |
|---|---|
| Base model | NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 |
| Base revision | a9904d24bcc1d289a1950fa9d2b978c47cf903b9 |
| Model size | 30B total parameters, 3B active |
| Adapter | LoRA, rank 8 |
| Teacher | DFS Large, deployment dfs-s46-span-opd-20260904-l |
| Objective | Completion-level reverse-KL distillation |
| Training cohort | 100 code-review prompts |
| Optimizer updates | 20 |
| Saved | October 2, 2026 |
Training recipe and provenance identify the teacher, source snapshots, tokenizer revision and checkpoint handling. The sampler and optimizer checkpoints were retained after the final update. This repository publishes the sampler export; it is not the optimizer checkpoint.
Current pilot results
A supplementary five-task Harbor pilot recorded 30% macro recall for Base Nemotron and 50% for this DFS Large update-20 checkpoint. OpenBabel improved from 0/1 to 1/1 known bugs matched after review; four tasks tied on recall. All five DFS20 trials used the exact sampler identified in this repository鈥檚 export manifest.
These are selected tasks with one trial per model/task. The paired task-bootstrap interval for the 20-point difference is 0 to 60 points and does not capture run-to-run uncertainty. No general quality or efficiency improvement is established. The runs used the native Tinker sampler; inference with the converted PEFT export has not been tested.
Both OpenBabel detectors initially flagged memory risks. Their final reports were converted and checked with GPT-6 Luna; only the trained report matched the known bug. The outcome applies to that full pipeline. Current pilot results and evidence.
What the linked traces show
The GitHub collection includes two current OpenBabel JSON traces and eight historical September 14 records. The historical trained runs used another checkpoint with a GLM teacher; they are not evaluations of these DFS weights. Current comparison 路 Historical comparison.
The JSON records preserve chronological event coverage, tool activity, source hashes and selected benign text, with labeled summaries and redactions. They omit private reasoning, source code and operational reproduction details; they are sanitized records, not complete raw Langfuse exports.
Access and use
Request access on this model page. An approved Hugging Face account can download the files after authenticating locally.
This is an adapter checkpoint, not a standalone copy of the base model. The converted PEFT adapter is 899,453,192 bytes (899 MB). Use the pinned base revision above and the format recorded in export-manifest.json. The original training rank is 8; the converter fuses some projections into rank 16. Local Transformers, PEFT and vLLM loading have not been validated for this release. Do not infer adapter compatibility from NVIDIA鈥檚 base-model serving instructions.
Use this research checkpoint for code-review evaluation in environments you control. Findings require review; a benchmark match does not prove a real-world exploit or establish production readiness.
License and attribution
NVIDIA developed the base model. Its pinned release uses the OpenMDW License Agreement, version 1.1. The unchanged upstream LICENSE preserves NVIDIA鈥檚 copyright notice and license text. depthfirst trained the adapter; this is a modified checkpoint, not an NVIDIA release.
Export verification
The exact sampler checkpoint was exported using the official Tinker cookbook adapter converter. All 12,010 tensor headers and 6,005 LoRA module pairs were checked against the pinned base-model shapes. No inference test was run. Nemotron expert LoRA serving is experimental in vLLM and is not supported by every serving engine.
| File | SHA-256 |
|---|---|
adapter_model.safetensors |
749cd0e2fa616f798fab177bf9df941f867f0dea387839f4ae56902ab9a4a400 |
adapter_config.json |
cebb27389a80d773efebe6625a1230606ba344f4ec383832012a06ebe232b5dd |