GPT-Redeaux

A GPT-2 XL fine-tune exported from Unsloth Studio. This revision contains complete BF16 model weights and F16 / Q8_0 GGUF exports. No separate base-model download or adapter merge is required.

Versions

Branch	Source run	Training steps	Saved (UTC)
main	1789180799	133	2026-09-12 02:42:58
run-1789180737-step-7	1789180737	7	2026-09-12 02:39:23

This branch is main, with 133 training steps and 7 epochs recorded in its checkpoint. These are separate training runs. The latest run is the default; no comparative quality evaluation was performed.

Architecture: GPT-2 XL (48 layers, 1,600 hidden dimensions, 25 attention heads), approximately 1.56 billion parameters, a 1,024-token context, and 50,258 vocabulary entries. The original GPT-2 tokenizer includes a training-added padding token.

Files

• model.safetensors: original, unquantized BF16 tensors, copied byte-for-byte from the completed run.
• config.json, generation_config.json, and tokenizer files: standalone Transformers loading configuration. Local machine references were removed and the inference KV cache enabled.
• gguf/GPT-Redeaux.F16.gguf: full floating-point GGUF converted from those weights.
• gguf/GPT-Redeaux.Q8_0.gguf: 8-bit quantized GGUF.
• export_info.json: source-run identifier, timestamps, weight checksum, and generation smoke test.
• SHA256SUMS: file checksums.

Transformers usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "4fhct4sd/GPT-Redeaux"
revision = "main"
# Authenticate with Hugging Face first if the repository is private.
tokenizer = AutoTokenizer.from_pretrained(repo, revision=revision)
model = AutoModelForCausalLM.from_pretrained(
    repo, revision=revision, dtype=torch.bfloat16, device_map="auto"
)
inputs = tokenizer("The meaning of life is", return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=64, do_sample=False,
                        pad_token_id=tokenizer.pad_token_id)
print(tokenizer.decode(output[0], skip_special_tokens=True))

This is a text-completion model; no chat template is supplied.

llama.cpp usage

Download either GGUF file from this branch, then run:

llama-completion -m GPT-Redeaux.Q8_0.gguf -p "The meaning of life is" -n 64 --no-conversation

Provenance and validation

The source run identifier is openai-community_gpt2-xl__project-----gpt2-cpt-ttext_1789125890__project----gpt2-cpt-ttext_1789130121__project---gpt2-cpt-ttext_1789139360__project--gpt2-cpt-rosetta-n-ttext_1789154027__project-gpt-ttext_rosetta-soulfft001_1789180799. This is a GPT-2 XL lineage with preceding local continued-training stages; the metadata above identifies the original architecture, not a claim that this run started directly from untouched upstream weights.

All 580 saved tensors were checked for finite values. The BF16 model was loaded and used for short greedy generation in Transformers. Both GGUF files were loaded and used for short generation in the installed llama.cpp build. These are export integrity checks, not benchmark results. Dataset composition and intended-use claims were not inferred from run names.