chris-linux-124m-gguf

Chris Linux 124M — GGUF

This repository contains GGUF releases of Chris Linux 124M, a 124,475,904-parameter GPT-2-style language model specialized for Linux terminal assistance and natural-language-to-command generation.

All GGUF files in this repository are derived from the same published model checkpoint:

linux_sft_core_v1/chris_linux_sft_epoch_08.pt

Canonical Transformers / SafeTensors release:

https://huggingface.co/christianrss/chris-linux-124m

Important: never automatically execute commands generated by this model. GGUF quantization changes the numerical representation of the weights and may also change individual outputs. Every quantized variant should be evaluated independently before use in an application.

Available files

The F16 file is the reference GGUF conversion. Q8_0 and Q4_0 are post-conversion quantizations of that F16 representation.

Model summary

Model lineage

Chris-GPT-2 124M
trained from scratch on 9,999,745,024 FineWeb-Edu tokens
        │
        ▼
FULL Linux continued pretraining — 10 epochs
219,807,744 token positions processed
        │
        ▼
Canonical Core SFT v1
3,335 train + 643 validation examples
106 intents
        │
        ▼
Core SFT epoch 08
validation loss 0.0389000066
        │
        ▼
Hugging Face SafeTensors conversion
        │
        ▼
F16 GGUF
        ├── Q8_0
        └── Q4_0

Why Core SFT

The model was specialized in stages. Linux continued pretraining improved its domain distribution, but CPT alone did not reliably turn natural-language instructions into precise shell commands. A larger heterogeneous instruction dataset also caused frequent confusion between nearby tools, flags, and argument patterns.

Core SFT v1 therefore follows the principle:

diverse inputs, canonical outputs

Examples:

show disk usage
→ df -h

show listening TCP ports
→ ss -lntp

show active TCP connections
→ ss -tnp

what process is using port 8080
→ sudo lsof -i :8080

show process with PID 8080
→ ps -p 8080

show traffic control configuration
→ tc qdisc show

find files larger than one gigabyte
→ find . -type f -size +1G

The Core SFT dataset contains 3,335 training examples, 643 validation examples, and 106 intents across 15 Linux-oriented categories.

Reference model evaluation

Before quantization, the Core SFT epoch-08 checkpoint obtained the following result on the Core command benchmark:

The 307-item Core benchmark contains 271 command examples, 28 diagnostic examples, and 8 safety examples. The exact-match result above refers specifically to the 271 command examples.

Important quantization note

The 66.79% figure belongs to the source Core SFT checkpoint/reference model evaluation. Do not automatically assume that F16, Q8_0, and Q4_0 all reproduce exactly the same score.

Quantization can change token rankings near decision boundaries. For publication-quality comparisons, run the same benchmark separately on each GGUF variant.

Prompt format

Use the same instruction format as the Transformers model:

User: <request>
Assistant:

Example:

User: show disk usage
Assistant:

Expected answer style:

df -h

The model is designed to generate short Linux-oriented answers rather than long conversational responses.

Running with Chris-Llama

Chris-Llama is the low-level inference runtime developed alongside the project.

Repository:

https://github.com/christianrss/chris-llama

Example:

./bin/chris_llama Chris-Linux-124M-Q8_0.gguf

Then enter a request in the interactive prompt, for example:

show disk usage

or:

show listening TCP ports

For highest-fidelity testing, start with:

./bin/chris_llama Chris-Linux-124M-F16.gguf

For a good quality/size tradeoff:

./bin/chris_llama Chris-Linux-124M-Q8_0.gguf

For a smaller local model:

./bin/chris_llama Chris-Linux-124M-Q4_0.gguf

The exact binary path may differ depending on how Chris-Llama was built.

GGUF conversion pipeline

The GGUF release is produced from the validated Hugging Face model rather than by directly rewriting the PyTorch checkpoint into GGUF.

Core SFT epoch-08 .pt
        │
        ▼
Hugging Face GPT2LMHeadModel
SafeTensors
        │
        ▼
llama.cpp convert_hf_to_gguf.py
        │
        ▼
Chris-Linux-124M-F16.gguf
        │
        ├── llama-quantize → Q8_0
        └── llama-quantize → Q4_0

This approach keeps the Transformers export as the canonical interoperable representation and uses the GGUF files as inference-oriented derivatives.

Tokenizer and padded vocabulary

The model uses the GPT-2 BPE tokenizer with 50,257 valid token IDs, while the trained embedding matrix contains 50,304 rows.

The extra rows were introduced for training efficiency and are not valid tokenizer outputs.

Inference runtimes should therefore restrict output sampling to the real GPT-2 token range:

0 ... 50256

and avoid sampling:

50257 ... 50303

The Transformers release handles this through generation_config.json. GGUF runtimes must preserve equivalent behavior in their sampling path.

Reproducibility identifiers

Source checkpoints recorded by the experiment:

Base Chris-GPT-2 checkpoint SHA-256:
2a62093c088046e05f73d72fda69ac17f3246cf71e4e5df1c4a01ffca135f67d

FULL Linux CPT checkpoint SHA-256:
cc7e2093e4198c5d88b93f1d3f6d152556b4a885d2095950145eeca7208ef5e7

Core SFT epoch-08 checkpoint SHA-256:
36c7143d75dcd9fc48a2fe8037f40ebf5834907890a94577d6510d2cb9075602

The repository should also contain SHA256SUMS.txt with checksums for the actual distributed GGUF files.

Verify a downloaded file with:

sha256sum Chris-Linux-124M-Q8_0.gguf

or verify all published files from the repository directory:

sha256sum -c SHA256SUMS.txt

Recommended variant

F16

Use F16 when:

• validating conversion quality;
• comparing against the Transformers/SafeTensors release;
• measuring quantization degradation;
• disk/RAM footprint is not the main constraint.

Q8_0

Q8_0 is the recommended default for this 124M model when local resources allow it. At this small parameter count, the memory savings of aggressive 4-bit quantization are less compelling than they are for multi-billion-parameter models, while preserving output stability is valuable for shell-command generation.

Q4_0

Use Q4_0 when minimizing model footprint matters more than preserving the closest possible behavior to the reference model. Validate command quality independently before integrating it into a terminal workflow.

Intended use

The GGUF release is intended for:

• local/offline Linux command-assistant experiments;
• CPU inference;
• lightweight terminal integrations;
• Chris-Llama runtime development;
• GGUF loader and quantization testing;
• benchmarking small specialized language models;
• educational systems and inference research.

Limitations

Chris Linux 124M is intentionally small. It does not have the broad reasoning or coding capacity of modern multi-billion-parameter assistants.

Possible failure modes include:

• choosing the wrong Linux utility;
• producing a valid utility with incorrect flags;
• confusing a port number with a PID or another numeric argument;
• composing mutually contradictory predicates;
• hallucinating Linux paths, files, flags, or tools;
• producing non-portable commands;
• failing on requests outside the curated Core intent space;
• failing to recognize a destructive operation when it is phrased in an unfamiliar way.

Quantization may introduce additional output differences.

Safety

Do not connect raw model output directly to command execution.

The model may generate commands that modify or destroy data. Any application built around this model should place a deterministic safety layer between generation and execution.

At minimum, consider blocking or requiring explicit confirmation for command families involving:

• rm with recursive/root-like targets;
• filesystem formatting tools such as mkfs;
• wipefs;
• raw block-device writes with dd;
• partitioning tools;
• recursive permission/ownership changes;
• shutdown/reboot operations;
• package removal;
• service changes;
• privileged operations through sudo.

A model-generated command being syntactically valid does not make it operationally safe.

Training-data provenance and licensing

The model combines several training stages:

• FineWeb-Edu-based base pretraining;
• Linux continued pretraining from multiple upstream documentation projects;
• manually curated Core SFT v1 examples.

The Linux documentation sources have heterogeneous upstream licenses. For that reason, this repository conservatively uses:

license: other

Users should review the provenance and licenses of upstream sources before redistribution or commercial use of derived weights.

Related repositories

• Transformers / SafeTensors model: https://huggingface.co/christianrss/chris-linux-124m
• Chris-GPT-2: https://github.com/christianrss/chris-gpt-2
• Chris-Llama: https://github.com/christianrss/chris-llama
• Chris-Torch: https://github.com/christianrss/chris-torch

Repository contents

A complete GGUF release is expected to look like:

README.md
Chris-Linux-124M-F16.gguf
Chris-Linux-124M-Q8_0.gguf
Chris-Linux-124M-Q4_0.gguf
publication_manifest.json
SHA256SUMS.txt
reproducibility/
    source_checkpoint.sha256
    core_benchmark_eval.json

Citation

Until the Linux-specialization paper has a stable public identifier, the project can be cited by its repository:

@misc{silva2026chrislinux,
  author       = {Christian Rafael de Souza Silva},
  title        = {Chris Linux 124M: A GPT-2-Style Language Model Specialized for Linux Terminal Assistance},
  year         = {2026},
  howpublished = {GitHub repository},
  url          = {https://github.com/christianrss/chris-gpt-2}
}

Disclaimer

Chris Linux 124M is an experimental research model. The author and contributors do not guarantee that generated shell commands are correct, safe, portable, or appropriate for any specific system.