Kestrel-34M-base

Base pretrained Looped Titans-MAG model (prelude + 3x4 looped core + coda, chunked neural memory + sliding-window attention, 16k byte-level BPE).

  • Repository: lillycham/kestrel
  • Stage: Pretraining (25,000 steps, ~819M tokens)
  • Parameters: 34,262,272 (~34.3M)
  • Architecture: Looped Titans-MAG (prelude=1, core=4 x 3 loops, coda=1, hidden_size=512, attention_window=256, memory_chunk_size=64, memory_heads=4, vocab_size=16384)
  • Final metrics: step 25000, train loss 3.263, val loss 3.230
  • Open SLM Eval: intelligence index 8.55 (raw 9.29) โ€” hellaswag: 26.6% (norm 25.7%), arc_easy: 42.1% (norm 39.6%), arc_challenge: 18.9% (norm 22.3%), piqa: 58.7% (norm 57.6%), arithmark3: 33.2% (norm 33.4%)

Usage

This model uses a custom KestrelForCausalLM architecture (trust_remote_code=True):

from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "lillycham/Kestrel-34M-base"
tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(repo_id, trust_remote_code=True)

prompt = tokenizer.apply_chat_template([{"role": "user", "content": "Hello!"}], tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt")
out = model.generate(inputs.input_ids, max_new_tokens=64, temperature=0.7, top_p=0.9)
print(tokenizer.decode(out[0], skip_special_tokens=False))

Training config summary

  • Steps: 25000, Batch size: 8 x grad_accum 4, Sequence length: 1024
  • Optimizers: Muon (lr=0.02) + AdamW (lr=0.001), weight decay 0.01
  • Precision: bf16 on cuda
Downloads last month
56
Safetensors
Model size
34.3M params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Collection including lillycham/Kestrel-34M-base