Kestrel-34M-anneal-base

Cooldown / annealing run initialized from Kestrel-34M-base with a mix including 20% synthetic step-by-step arithmetic.

  • Repository: lillycham/kestrel
  • Stage: Annealing (5,000 steps from Kestrel-34M-base, with 20% synthetic arithmetic)
  • Initialized from: lillycham/Kestrel-34M-base
  • 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 5000, train loss 1.776, val loss 2.045
  • Open SLM Eval: intelligence index 8.78 (raw 9.36) โ€” hellaswag: 26.3% (norm 26.3%), arc_easy: 41.7% (norm 38.7%), arc_challenge: 19.5% (norm 22.2%), piqa: 58.1% (norm 57.0%), arithmark3: 35.3% (norm 35.5%)

Usage

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

from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "lillycham/Kestrel-34M-anneal-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: 5000, Batch size: 8 x grad_accum 4, Sequence length: 1024
  • Optimizers: Muon (lr=0.015) + AdamW (lr=0.00075), weight decay 0.01
  • Precision: bf16 on cuda
Downloads last month
67
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-anneal-base