erebus-v2-1.5b-tool

Erebus v2 1.5B - Tool

A 1.5B parameter function-calling model fine-tuned from erebus-v2-1.5b-base on Salesforce xLAM-60k for structured tool use.

Training

Findings

The model correctly selects the right tool from a schema and produces JSON function calls. Example outputs:

• get_weather with {"location": "Paris"} - correct tool and argument
• calculator with {"expression": "25 * 47"} - chose calculator over search
• send_email with {"subject": "Hello"} - correct tool selection

Known Limitations

• Schema leakage: Arguments sometimes include schema-like fields ("type": "object", "properties") instead of flat values.
• Missing/wrong args: Occasionally drops required arguments or fills in placeholder values.
• Post-JSON degeneration: After the initial JSON object, output degenerates into repetition. Use a stop token or extract the first JSON object.
• Same base model limitations: Small pretraining budget (5.5B tokens) limits overall language understanding.

Best used with post-processing to extract the first JSON object from the output.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import json

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

messages = [
    {"role": "system", "content": 'You have access to: {"name": "get_weather", "parameters": {"properties": {"location": {"type": "string"}}}}'},
    {"role": "user", "content": "What's the weather in Tokyo?"},
]
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=100)
response = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
# Extract first JSON object from response
print(response.split("\n")[0])

Variants

License

Apache 2.0