snehitsh

qwen3-1.7b-toolcall-lora-v1

README

License: apache-2.0

Quick start

python

import json, torchfrom huggingface_hub import hf_hub_downloadfrom peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base_id, adapter_id = "Qwen/Qwen3-1.7B", "snehitsh/qwen3-1.7b-toolcall-lora-v1"tok = AutoTokenizer.from_pretrained(base_id)model = AutoModelForCausalLM.from_pretrained(base_id, dtype=torch.bfloat16, device_map="cuda")  # older transformers: torch_dtype=model = PeftModel.from_pretrained(model, adapter_id).eval()tools = json.load(open(hf_hub_download(adapter_id, "tools.json")))   # the five tool schemas messages = [    {"role": "system", "content": "You are a helpful assistant. The current date and time is 2026-02-17 14:20 (Tuesday)."},    {"role": "user", "content": "remind me tomorrow at 7pm to cook dinner"},]prompt = tok.apply_chat_template(messages, tools=tools, add_generation_prompt=True, tokenize=False,                                 enable_thinking=False)       # trained in non-thinking mode onlyinputs = tok(prompt, return_tensors="pt").to("cuda")out = model.generate(**inputs, max_new_tokens=256, do_sample=False)print(tok.decode(out[0, inputs.input_ids.shape[1]:], skip_special_tokens=False))# <tool_call># {"name": "set_reminder", "arguments": {"message": "cook dinner", "time": "2026-02-18 19:00"}}# </tool_call><|im_end|>

Run the returned call yourself, then append the assistant message and a tool message with the result and generate again to get a plain-text answer. Always validate a call before executing it.

Conventions the model learned

Reminders need the current date. Put The current date and time is YYYY-MM-DD HH:MM (Weekday). in the system message. Reminder time is YYYY-MM-DD HH:MM.
Units use short names: km mi m ft in cm yd kg lb g oz c f k l gal ml cup mph kph.
calculate takes plain arithmetic (+ - * / ** and parentheses).
get_weather omits unit unless the user asks for one.
enable_thinking=False must be used when rendering the chat template.

Results

Held-out test set of 185 examples written for this project (frozen before training; greedy decoding; same harness for both columns):

Table with columns: Category (n), Base Qwen3-1.7B, + this adapter
Category (n)	Base Qwen3-1.7B	+ this adapter
Overall (185)	68.7%	95.1%
calculate (15)	80.0%	100%
get_weather (15)	60.0%	100%
get_datetime (15)	80.0%	100%
convert_units (15)	100%	100%
set_reminder (15)	46.7%

Valid-JSON tool calls: 82% to 100%. False tool calls where none was correct: 6.7% to 0%.

A separate spot check on 31 new prompts written after training, with different phrasing, passed 29 of 31. Both misses were date arithmetic.

Training

LoRA rank 16, alpha 32, dropout 0.05 on q/k/v/o/gate/up/down projections in all 28 layers (17.4M trainable parameters, 69.8 MB).
2 epochs, 508 steps, effective batch 16, learning rate 2e-4 (20 warmup steps, cosine decay), bf16, one 8 GB laptop GPU. Loss computed on the assistant's answer tokens only.
Data (about 4,000 examples): about 1,600 synthetic examples for the five tools with programmatically computed labels (not released), 2,000 examples from the xLAM-60k dataset, and 400 no-tool examples. Prompts were up to about 1.4k tokens long.

Limitations

This is v1. Five tools only, one training run per setting, no hyperparameter search. Expect rough edges.
Date and time arithmetic is the weak spot. Mistakes seen: adding minutes ("in 45 minutes" from 14:20 gave 14:55), weekday offsets (picked the wrong week or day), and AM/PM for phrases like "tonight at 9". Check reminder times before relying on them.
The test set is in-distribution. It was written in the same style as the training data, with near-duplicates removed. It has 185 examples and was run once per model, so real-world accuracy will likely be lower than 95%.
English only, non-thinking mode only, and no testing on long conversations (training prompts were up to about 1.4k tokens).
Only these five tools are supported. New tools would need new data. For requests that need an unavailable tool, the model may reply that none of its tools can help.
Not evaluated for safety or for regressions on general abilities beyond plain chat. Outputs can be wrong: validate tool calls before executing them.

License and attribution

This adapter is released under Apache 2.0, matching the base model Qwen3-1.7B (Apache 2.0).
Training data included a subset of Salesforce/xlam-function-calling-60k (CC-BY-4.0). Credit: Salesforce, APIGen: Automated Pipeline for Generating Verifiable and Diverse Function-Calling Datasets (arXiv:2406.18518).

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

snehitsh

Model Tree

Base

Qwen/Qwen3-1.7B

Adapter

this model

Input Modalities

Text

Output Modalities