🦉 Owl-v1

1. Model Intro

Owl-v1 is a lightweight 10.8 M parameter language model built from scratch for learning and integration purposes.

2. What Can It Generate?

Character name / scene opener prompts.

Prompt	Output	Quality
"HAMLET:\n"	Character dialogue in Early Modern English	⭐⭐⭐
"Enter KING and QUEEN.\n"	Stage directions + speech	⭐⭐
"To be, or not to be"	Probabilistic monologue continuation	⭐⭐
"Shall I compare thee"	Poetic continuation (metre may drift)	⭐
"\n" (cold start)	Free-form dramatic scene	⭐⭐

3. Usage

Requires trust_remote_code=True — CharTokenizer and ModelArchitecture are custom classes.

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

tokenizer = AutoTokenizer.from_pretrained("IstishadAlamTishad/Owl-v1", trust_remote_code=True)
model     = AutoModelForCausalLM.from_pretrained("IstishadAlamTishad/Owl-v1", trust_remote_code=True)
model.eval()

ids = tokenizer.encode("HAMLET:\n", return_tensors="pt")
with torch.no_grad():
    out = model.generate(ids, max_new_tokens=200, do_sample=True, temperature=0.8, top_k=40)
print(tokenizer.decode(out[0].tolist()))

4. Model Training & Specs

Dataset	Tokens	Steps	Batch
TinyShakespeare	1,115,394	5,000	64

Property	Value
Parameters	10.8 M
Layers / Heads / D_embd	6 / 6 / 384
Context length	256 tokens
Vocab	65 chars

5. Performance Metrics Curves

Fig 1 — Perplexity: Train PPL → ~1, Val PPL bottoms at ~5 (step ~800) then climbs to ~50 — clear overfitting on the small corpus.	Fig 2 — Throughput: Stable ~37–38 k tokens/sec after a brief warm-up spike.

Built for learning and integration