---
language:
  - en
license: apache-2.0
tags:
  - gpt2
  - text-generation
  - synthara
  - deprecated
  - gguf
  - llama-cpp
pipeline_tag: text-generation
library_name: gguf
inference:
  parameters:
    max_new_tokens: 200
    temperature: 0.9
    top_p: 0.95
    repetition_penalty: 1.1
widget:
  - text: "Once upon a time"
    example_title: "Story start"
  - text: "The quick brown fox"
    example_title: "Sentence completion"
providers:
  - hf-inference
---

# Synthara Legacy

> **Deprecated.** This model is no longer maintained and is not recommended
> for any production or serious research use. It exists purely as a historical
> artifact for the Tripplet Research organisation. See newer Synthara releases
> if any are available.

---

**Honest disclaimer:** Synthara Legacy is not a good model. It was built as an
early proof-of-concept with randomly initialised weights and no fine-tuning on
meaningful data. Output quality is poor — expect incoherent or repetitive text.
It is published here for transparency and archival purposes only.

---

## Architecture

| Property | Value |
|---|---|
| Base architecture | GPT-2 |
| Parameters | ~51.5 M |
| Layers | 8 |
| Attention heads | 8 |
| Embedding dim | 512 |
| Context length | 1 024 tokens |
| Tokenizer | GPT-2 fast (Apache 2.0) |

## Status

**DEPRECATED — do not use in production.**

This checkpoint has never been trained on any dataset. Weights are random
initialisations only. It will not produce useful output without significant
fine-tuning.

## License

Apache 2.0 — see [LICENSE](https://www.apache.org/licenses/LICENSE-2.0).

## Credits

- Architecture based on the open GPT-2 specification (OpenAI, MIT licence).
- Tokenizer from `openai-community/gpt2` (MIT licence).
- Built with [Transformers](https://github.com/huggingface/transformers) (Apache 2.0).
- Published by [Tripplet Research](https://huggingface.co/tripplet-research).

This model is **not** derived from any unlicensed third-party checkpoint.