Irfanuruchi/Nanbeige4.2-3B-CE-v1.0

Nanbeige4.2-3B-CE v1.0

Nanbeige4.2-3B-CE is a standalone Computer Engineering and systems-focused fine-tune of Nanbeige/Nanbeige4.2-3B.

This repository contains the fully merged BF16 model. No PEFT adapter is required for inference.

Release

• Version: v1.0
• Base: Nanbeige/Nanbeige4.2-3B
• Base revision: 3384e426066d1a49c3aea90a7190b81260a6533f
• Source training checkpoint: v1.0.5
• Format: BF16 Safetensors
• Architecture: NanbeigeForCausalLM
• Tokenizer: LlamaTokenizer

The final model was produced by loading the pinned base in FP32, applying the frozen FP32 LoRA adapter, merging in FP32, and casting the completed model once to BF16.

Focus areas

The model is specialized toward:

• operating systems and Linux
• computer architecture
• virtual memory and page faults
• caching and memory hierarchy
• concurrency and atomics
• networking and DNS
• Docker and infrastructure
• NUMA and multithreaded performance
• embedded and systems engineering

Integrity

Canonical BF16 weight shards:

872a9a53bdd9aec68730c78f386d6a61a1db7d28901fe54064cf560c0a296b3b model-00001-of-00002.safetensors

45043aa026e33f03804f3a92bb9ad38cd68836beeab766d6741c0699922956c1 model-00002-of-00002.safetensors

A complete SHA256SUMS manifest is included.

Usage

Install PyTorch and Transformers, then load the model with trust_remote_code=True.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL = "Irfanuruchi/Nanbeige4.2-3B-CE-v1.0"

tokenizer = AutoTokenizer.from_pretrained(
    MODEL,
    trust_remote_code=True,
    use_fast=False,
)

model = AutoModelForCausalLM.from_pretrained(
    MODEL,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [{
    "role": "user",
    "content": "Why can false sharing reduce multithreaded performance?"
}]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    enable_thinking=False,
    return_tensors="pt",
).to(model.get_input_embeddings().weight.device)

with torch.inference_mode():
    output = model.generate(
        inputs,
        max_new_tokens=160,
        do_sample=False,
    )

print(tokenizer.decode(
    output[0][inputs.shape[-1]:],
    skip_special_tokens=True,
))

Release validation used enable_thinking=False and deterministic decoding.

Training

The selected checkpoint was produced using QLoRA/LoRA.

• LoRA rank: 16
• LoRA alpha: 32
• LoRA dropout: 0.05
• 4-bit NF4 training base
• BF16 compute
• double quantization enabled
• approximately 24M trainable adapter parameters

Target modules:

q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

Multiple later targeted SFT, completion-only, checkpoint-sweep, and preference-optimization experiments were evaluated. None produced a consistently better overall checkpoint without regressions, so v1.0.5 remained the selected source checkpoint.

Validation

The release was evaluated with a frozen 28-prompt systems regression suite covering:

• x86-64 translation and TLB behavior
• page faults and copy-on-write
• C++ memory ordering
• sequential consistency
• caches and coherence
• false sharing
• Docker DNS
• DNS UDP/TCP behavior
• Linux load average
• worker contention
• NUMA placement

The merged standalone artifact was also verified to load locally without requiring the original base-model repository.

Known limitations

This model is not claimed to be factually perfect.

Known difficult cases include:

• omission of std::memory_order_release when enumerating every standard C++ memory order
• imprecise wording around some major/minor page-fault edge cases
• incorrect descriptions of Docker 127.0.0.11 on some prompts
• occasionally weak explanations of cache-line ping-pong
• precision-sensitive terminology around TLB filling, cache policies, and sequential consistency

Inside a container attached to a user-defined Docker network, 127.0.0.11 is Docker's embedded DNS resolver.

For major versus minor page faults, the key distinction is whether resolving the fault requires backing-store I/O.

Independent verification is recommended for production-critical engineering decisions.

Base model

Built from Nanbeige/Nanbeige4.2-3B.

Pinned revision:

3384e426066d1a49c3aea90a7190b81260a6533f

The upstream Nanbeige model is distributed under Apache-2.0.

Other formats

Additional releases are planned for:

• GGUF
• MLX
• OpenVINO
• original LoRA adapter