Talos Mini 1M — OASST1

Experimental 1,000,320-parameter Talos Mini language model trained on a small subset of OpenAssistant/oasst1.

This repository contains the native Talos checkpoint and native Talos Byte-Level BPE tokenizer.

Model specifications

  • Parameters: 1,000,320
  • Architecture: TalosGPT decoder-only transformer
  • Hidden size: 128
  • Layers: 3
  • Attention heads: 8
  • KV heads: 4
  • Head dimension: 16
  • Dense FFN intermediate size: 512
  • Vocabulary size: 1,024
  • Maximum sequence length: 512
  • Tokenizer merges: 764
  • Training data: 1,800 OASST1 messages
  • Evaluation data: 200 held-out OASST1 messages
  • Training sequence length: 64
  • Training batch size: 4
  • Optimizer: AdamW
  • Learning rate: 3e-3

Training

The model was trained using the existing Talos training stack.

Published checkpoint:

step-26925.pt

The checkpoint contains the native Talos model state and training metadata.

Results

At step 26,925:

  • Validation loss: approximately 2.254
  • Validation perplexity: approximately 9.52
  • Validation next-token accuracy: approximately 35%
  • Evaluated tokens: 102,816

These results come from a small research experiment and are not a benchmark of general language-model capability.

The model can generate recognizable language patterns, but it is not a coherent general-purpose conversational model.

Tokenizer

The tokenizer is the native Talos ByteLevelBPETokenizer.

It is not a Hugging Face GPT-2 tokenizer conversion.

The tokenizer uses Talos-native byte IDs and merge IDs.

  • Vocabulary: 1,024
  • Merges: 764

Status

Demonstrated

  • 1M-parameter Talos model construction
  • Native Talos tokenizer
  • OASST1 data loading
  • Train/evaluation split
  • Training
  • Checkpoint saving/loading
  • Held-out evaluation
  • Text generation
  • KV-cache equivalence testing
  • Greedy generation

Limitations

This is an experimental dev-scale model trained on a small dataset and short sequence length.

It is not intended to compete with modern general-purpose language models.

The 1M model is a scaling experiment between the 254K Talos Mini prototype and larger future Talos configurations.

Long-term Talos scaling targets are research goals and are not represented here as demonstrated capabilities.

Downloads last month
119
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Dataset used to train PrometheanStudio/talos-mini-1m-oasst1