rvindra

nemotron-3.5-lightning-bpf-guardian

Visual Performance Progression

BPF-Guardian Benchmark Performance Across Training Stages

4-Panel Detailed Suite Breakdown

Domain Robustness and Complexity Scaling

Performance Highlights (Live Linux Kernel 6.8 VPS Execution)

Table with columns: Evaluation Benchmark, Benchmark Size, Base Nemotron 30B, Prior Baseline (Qwen3-8B SFT v2), Nemotron BPF-Guardian SFT v1, Nemotron BPF-Guardian RL (Solve@2), Relative Improvement
Evaluation Benchmark	Benchmark Size	Base Nemotron 30B	Prior Baseline (Qwen3-8B SFT v2)	Nemotron BPF-Guardian SFT v1	Nemotron BPF-Guardian RL (Solve@2)	Relative Improvement
Protected Private Synthesis	120 tasks	0 / 120 (0.0%)	31 / 120 (25.8%)	54 / 120 (45.0%)	59 / 120 (49.2%)	+90.3%
Protected Standalone Repair	120 tasks	79 / 120 (65.8%)	85 / 120 (70.8%)	91 / 120 (75.8%)	91 / 120 (75.8%)	+7.1%
Confirmation Suite	60 tasks	20 / 60 (33.3%)	33 / 60 (55.0%)	42 / 60 (70.0%)	44 / 60 (73.3%)	+33.3%
N3 Stratified Dev Suite	48 tasks	—	18 / 48 (37.5%)	24 / 48 (50.0%)	23 / 48 (47.9%)	+27.7%
Combined Protected Suite	276 tasks	79 / 276 (28.6%)	137 / 276 (49.6%)	168 / 276 (60.9%)	172 / 276 (62.3%)	+25.6%

All results verified with zero mock verifiers using clang-18 -target bpf -O2 and kernel execution on Linux 6.8.0-106-generic. Paired McNemar test on combined suite confirms statistical significance ( $p = 1.38 \times 1 0 - 5$ , $p < 0.0001$ ).

Available Checkpoints & Branches

main: Nemotron SFT v1 checkpoint (optimized for standalone synthesis and repair, 75.8% repair pass rate, 168/276 on combined benchmark).
rl-n3: Nemotron RL N3 checkpoint (optimized for two-turn interactive repair with compiler/verifier feedback, 49.2% Solve@2 on protected synthesis, 73.3% on confirmation).

Live In-Kernel Verification Pipeline

Every generated XDP program undergoes strict 4-stage validation:

markdown
[Candidate C Code]
       │
       ▼
1. Structural Lint & Sanitization (SEC("xdp") header, helper verification)
       │
       ▼
2. Compilation (clang-18 -target bpf -O2 -g -Wall -Werror)
       │
       ▼
3. In-Kernel Verifier Check (bpftool prog load into Linux 6.8.0-106-generic)
       │
       ▼
4. Dynamic Packet Execution (BPF_PROG_TEST_RUN socket test fixtures)
       │
       ├── PASS: Reward = 1.0 (Pass@1)
       └── FAIL: Extract standardized compiler / verifier / assertion diff diagnostic
             │
             ▼
[Turn 2: Repaired Code] ──► Re-run Stages 1-4 ──► PASS: Reward = 0.9 (Solve@2)

Published Datasets

rvindra/bpf-guardian-sft: Combined instruction-tuning corpus merging v1 and v2 synthesis and repair datasets (2,320 examples: 1,913 train / 407 val). Strictly audited with 0% calibration benchmark overlap.
rvindra/bpf-guardian-rl: Certified 264-task RLVR benchmark with raw packet hex fixtures and expected XDP verdicts. Packaged with formal contamination_audit.json.

Functional Capabilities

The model synthesizes pure, self-contained, and kernel-verified C code across four core network programming domains:

Packet Filtering & Security (pfs): DDoS defenses, port knocking state machines, token bucket policers, bloom filters.
Network Routing & Forwarding (nrf): Maglev consistent hashing, ECMP multipath routers, LPM trie routing, proxy ARP, tunnel encapsulation.
Packet Inspection & Telemetry (pit): Heavy hitters (Count-Min Sketch), flow telemetry, TCP RTT and options analysis.
Packet Transformation & Rewrite (ptr): VLAN/QinQ manipulation, GTP-U/VXLAN decapsulation, IPv4/IPv6 header rewrites.

Usage with PEFT & Transformers

python
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base_model_id = "nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16"
peft_model_id = "rvindra/nemotron-3.5-lightning-bpf-guardian"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Load SFT model (revision="main") or RL model (revision="rl-n3")
model = PeftModel.from_pretrained(base_model, peft_model_id, revision="rl-n3")

prompt = \"\"\"You are an expert Linux kernel eBPF developer. Write a complete, self-contained XDP C program that inspects incoming IPv4 TCP packets, extracts the destination port, and drops packets targeting port 8080.
Output ONLY the raw C source code. Do not wrap with markdown code fences.\"\"\"

messages = [
    {"role": "system", "content": "You are an expert Linux kernel eBPF developer. Output ONLY valid, compilable C code."},
    {"role": "user", "content": prompt}
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to("cuda")
outputs = model.generate(inputs, max_new_tokens=2048, do_sample=False)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))

Training Methodology

Base Architecture: Nemotron-3.5-Lightning (30B Total Parameters, 3.5B Active Parameters per token MoE).
LoRA Configuration: Rank 32, Alpha 64, Target Modules: all attention and MLP projections (q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj).
SFT Stage: 3 epochs on 1,600 verified instruction-tuning examples with family-heldout cross-validation.
RL Stage: 30 steps of Importance Sampling RLVR with live Linux kernel verifier rewards and $β = 0.05$ KL regularization.

Notes on AI

This project was primarily conducted using AI:

Model training
Dataset generation
Scripting and coding

Acknowledgements

Compute: Fine-tuned and evaluated via Thinking Machines.

Citation

bibtex
@misc{nemotron_bpf_guardian_2026,
  author = {Tarunokusumo, Ravindra},
  title = {Nemotron-3.5-Lightning BPF-Guardian: Verified In-Kernel eBPF/XDP Generation},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/rvindra/nemotron-3.5-lightning-bpf-guardian}}
}

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

rvindra

Model Tree

Base

nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16

Adapter

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

Visual Performance Progression

BPF-Guardian Benchmark Performance Across Training Stages

4-Panel Detailed Suite Breakdown

Domain Robustness and Complexity Scaling

Performance Highlights (Live Linux Kernel 6.8 VPS Execution)

Table with columns: Evaluation Benchmark, Benchmark Size, Base Nemotron 30B, Prior Baseline (Qwen3-8B SFT v2), Nemotron BPF-Guardian SFT v1, Nemotron BPF-Guardian RL (Solve@2), Relative Improvement
Evaluation Benchmark	Benchmark Size	Base Nemotron 30B	Prior Baseline (Qwen3-8B SFT v2)	Nemotron BPF-Guardian SFT v1	Nemotron BPF-Guardian RL (Solve@2)	Relative Improvement
Protected Private Synthesis	120 tasks	0 / 120 (0.0%)	31 / 120 (25.8%)	54 / 120 (45.0%)	59 / 120 (49.2%)	+90.3%
Protected Standalone Repair	120 tasks	79 / 120 (65.8%)	85 / 120 (70.8%)	91 / 120 (75.8%)	91 / 120 (75.8%)	+7.1%
Confirmation Suite	60 tasks	20 / 60 (33.3%)	33 / 60 (55.0%)	42 / 60 (70.0%)	44 / 60 (73.3%)	+33.3%
N3 Stratified Dev Suite	48 tasks	—	18 / 48 (37.5%)	24 / 48 (50.0%)	23 / 48 (47.9%)	+27.7%
Combined Protected Suite	276 tasks	79 / 276 (28.6%)	137 / 276 (49.6%)	168 / 276 (60.9%)	172 / 276 (62.3%)	+25.6%

Available Checkpoints & Branches

main: Nemotron SFT v1 checkpoint (optimized for standalone synthesis and repair, 75.8% repair pass rate, 168/276 on combined benchmark).
rl-n3: Nemotron RL N3 checkpoint (optimized for two-turn interactive repair with compiler/verifier feedback, 49.2% Solve@2 on protected synthesis, 73.3% on confirmation).

Live In-Kernel Verification Pipeline

Every generated XDP program undergoes strict 4-stage validation:

markdown
[Candidate C Code]
       │
       ▼
1. Structural Lint & Sanitization (SEC("xdp") header, helper verification)
       │
       ▼
2. Compilation (clang-18 -target bpf -O2 -g -Wall -Werror)
       │
       ▼
3. In-Kernel Verifier Check (bpftool prog load into Linux 6.8.0-106-generic)
       │
       ▼
4. Dynamic Packet Execution (BPF_PROG_TEST_RUN socket test fixtures)
       │
       ├── PASS: Reward = 1.0 (Pass@1)
       └── FAIL: Extract standardized compiler / verifier / assertion diff diagnostic
             │
             ▼
[Turn 2: Repaired Code] ──► Re-run Stages 1-4 ──► PASS: Reward = 0.9 (Solve@2)

Published Datasets

rvindra/bpf-guardian-sft: Combined instruction-tuning corpus merging v1 and v2 synthesis and repair datasets (2,320 examples: 1,913 train / 407 val). Strictly audited with 0% calibration benchmark overlap.
rvindra/bpf-guardian-rl: Certified 264-task RLVR benchmark with raw packet hex fixtures and expected XDP verdicts. Packaged with formal contamination_audit.json.

Functional Capabilities

The model synthesizes pure, self-contained, and kernel-verified C code across four core network programming domains:

Packet Filtering & Security (pfs): DDoS defenses, port knocking state machines, token bucket policers, bloom filters.
Network Routing & Forwarding (nrf): Maglev consistent hashing, ECMP multipath routers, LPM trie routing, proxy ARP, tunnel encapsulation.
Packet Inspection & Telemetry (pit): Heavy hitters (Count-Min Sketch), flow telemetry, TCP RTT and options analysis.
Packet Transformation & Rewrite (ptr): VLAN/QinQ manipulation, GTP-U/VXLAN decapsulation, IPv4/IPv6 header rewrites.

Usage with PEFT & Transformers

python
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base_model_id = "nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16"
peft_model_id = "rvindra/nemotron-3.5-lightning-bpf-guardian"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Load SFT model (revision="main") or RL model (revision="rl-n3")
model = PeftModel.from_pretrained(base_model, peft_model_id, revision="rl-n3")

prompt = \"\"\"You are an expert Linux kernel eBPF developer. Write a complete, self-contained XDP C program that inspects incoming IPv4 TCP packets, extracts the destination port, and drops packets targeting port 8080.
Output ONLY the raw C source code. Do not wrap with markdown code fences.\"\"\"

messages = [
    {"role": "system", "content": "You are an expert Linux kernel eBPF developer. Output ONLY valid, compilable C code."},
    {"role": "user", "content": prompt}
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to("cuda")
outputs = model.generate(inputs, max_new_tokens=2048, do_sample=False)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))

Training Methodology

Base Architecture: Nemotron-3.5-Lightning (30B Total Parameters, 3.5B Active Parameters per token MoE).
LoRA Configuration: Rank 32, Alpha 64, Target Modules: all attention and MLP projections (q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj).
SFT Stage: 3 epochs on 1,600 verified instruction-tuning examples with family-heldout cross-validation.
RL Stage: 30 steps of Importance Sampling RLVR with live Linux kernel verifier rewards and $β = 0.05$ KL regularization.

Notes on AI

This project was primarily conducted using AI:

Model training
Dataset generation
Scripting and coding

Acknowledgements

Compute: Fine-tuned and evaluated via Thinking Machines.

Citation

bibtex
@misc{nemotron_bpf_guardian_2026,
  author = {Tarunokusumo, Ravindra},
  title = {Nemotron-3.5-Lightning BPF-Guardian: Verified In-Kernel eBPF/XDP Generation},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/rvindra/nemotron-3.5-lightning-bpf-guardian}}
}

nemotron-3.5-lightning-bpf-guardian

README

Visual Performance Progression

Performance Highlights (Live Linux Kernel 6.8 VPS Execution)

Available Checkpoints & Branches

Live In-Kernel Verification Pipeline

Published Datasets

Functional Capabilities

Usage with PEFT & Transformers

Training Methodology

Notes on AI

Acknowledgements

Citation

Explore FriendliAI today

README

Visual Performance Progression

Performance Highlights (Live Linux Kernel 6.8 VPS Execution)

Available Checkpoints & Branches

Live In-Kernel Verification Pipeline

Published Datasets

Functional Capabilities

Usage with PEFT & Transformers

Training Methodology

Notes on AI

Acknowledgements

Citation