Cosmos-T-80M Demo

Chain-of-Thought Chat Demo

79.7M params 12 attention layers Qwen2.5 tokenizer Apache-2.0

Pretrained from scratch on wop/XXXXXL-chain-of-thought · Model card: wop/Cosmos-T-80M

⚠️ Research / demo model. Only 840 training conversations, so the model is heavily overfit and will hallucinate confidently outside its training distribution. Treat it as a stylish parrot — not a fact source.

Cosmos-T-80M

Examples

	System prompt	Temperature	Top-K	Context window (max 1028)	Max new tokens (max 1028)

System prompt

Controls the model's reasoning mode.

Temperature

0 = greedy (deterministic). Low values keep the model on-rails.

0 2

Top-K

Sample from the K most likely tokens. K=1 = pure argmax.

1 200

Context window (max 1028)

How many input tokens the model sees.

64 1028

Max new tokens (max 1028)

Hard cap on response length.

16 1028

**Tips** — Keep `temp = 0.1` and `top_k = 1` for the most coherent output. Crank `temp` up to 0.8+ for more creative (but messier) replies. Clear the chat if responses start looping.