
Base model	erebus-v2-1.5b-base (5.5B token pretrain)
SFT dataset	HuggingFaceTB/smoltalk (~1M examples)
Epochs	1
LR	2e-5 (cosine decay)
Batch	4 per device x 4 GPUs x 4 grad accum = 64
Steps	16,245
Time	34.5 hours on 4x A100-SXM4-80GB
Final loss	~1.85

Known Limitations

Repetition: The model tends to repeat phrases and sentences, especially in longer outputs. Greedy decoding amplifies this. Using repetition_penalty=1.2 helps.
No stop control: The model often doesn't know when to stop generating, producing verbose responses that loop.
Weak reasoning: GSM8K score is ~1.4% (flexible) / 0.08% (strict). Multi-step math reasoning is essentially absent due to the small pretraining budget (5.5B tokens vs 18T for Qwen2.5).
Code: Can produce simple correct functions (e.g. is_prime) but explanations degenerate into repetition.

These are primarily a consequence of the small pretraining token budget, not the architecture or SFT data.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "soyrsoyr/erebus-v2-1.5b-instruct",
    torch_dtype="bfloat16",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("soyrsoyr/erebus-v2-1.5b-instruct")

messages = [{"role": "user", "content": "What is the capital of France?"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=200, repetition_penalty=1.2)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Variants

Variant	Description	Link
Base	Pretrained model	soyrsoyr/erebus-v2-1.5b-base
Instruct	SFT on SmolTalk (this)	soyrsoyr/erebus-v2-1.5b-instruct
Tool	SFT on xLAM for function calling	soyrsoyr/erebus-v2-1.5b-tool

License

Apache 2.0