CodeXomics

CodeXomics-ToolAgent-4B-v1

README

License: apache-2.0

Fine-tuning

Method: QLoRA (rank 16, scale 32.0, dropout 0.05, 4 layers) with MLX-LM 0.31.2 / MLX 0.32.0
Trainable parameters: 4.058M (0.096%)
Optimizer: AdamW, learning rate 1.0e-5, effective batch size 4, 200 iterations
Maximum sequence length: 3,072 tokens; prompt masking enabled
Hardware: Apple M3 Max; peak memory 191 GB (including swap)
Checkpoint selection: iteration 75 (validation loss 0.020); test loss 0.074 (perplexity 1.077)
Training data: CodeXomics-ToolCalling-v1 (373/123/30 train/validation/test examples)

Evaluation

On the CodeXomics Benchmark (172 automatic tests: 143 single-operation, 29 multi-step), evaluated in the real application loop with task-completion scoring plus execution evidence. Both the fine-tuned model and the un-fine-tuned Qwen3.5-4B baseline were evaluated in three independent complete sessions; results were identical across sessions for both models.

Table with columns: Suite, Qwen3.5-4B (base), CodeXomics-ToolAgent-4B-v1
Suite	Qwen3.5-4B (base)	CodeXomics-ToolAgent-4B-v1
Simple	139/143	143/143
Complex	26/29	29/29
Total	165/172 (95.9%)	172/172 (100%)

Fine-tuning improved the overall accuracy by 7 tests (+4 simple, +3 complex). Inference settings: temperature 0, thinking enabled.

Inference speed (mean ± SD over three independent runs; offline harness, 172 tests, Ollama Q4_K_M on Apple M3 Max):

Table with columns: Metric, Qwen3.5-4B (base), CodeXomics-ToolAgent-4B-v1
Metric	Qwen3.5-4B (base)	CodeXomics-ToolAgent-4B-v1
Average latency per test (s)	12.0 ± 0.1	10.7 ± 0.9
Simple suite (s)	9.2 ± 0.1	8.6 ± 0.7
Complex suite (s)	25.8 ± 0.3	21.5 ± 1.9
Average latency per tool call (s)	7.9 ± 0.1	7.3 ± 0.6
Generation throughput (tokens/s)	33.9 ± 0.4	33.9 ± 2.8
Generated tokens per test

CodeXomics Benchmark: base vs. fine-tuned 4B

Figure 1. Two-panel comparison between the un-fine-tuned Qwen3.5-4B baseline and CodeXomics-ToolAgent-4B-v1. Left: CodeXomics Benchmark pass rates (simple/complex/total; three sessions each, identical results). Right: average inference latency per test with error bars (mean ± SD over three runs; Ollama Q4_K_M, Apple M3 Max). Latency values are the mean of the per-suite rows above; error bars show the run-to-run SD.

Usage

python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("CodeXomics-ToolAgent-4B-v1")
tokenizer = AutoTokenizer.from_pretrained("CodeXomics-ToolAgent-4B-v1")

For deployment in CodeXomics, the model is served through Ollama as qwen3.5:4b-codexomics-tools-v5 (Q4_K_M, 2.7 GB) with native tool calling and thinking enabled.

Limitations

The model is specialized for CodeXomics genomic workflows; generalization to other tool-calling domains was not evaluated.

Citation

bibtex
@software{codexomics-toolagent-v1,
  title = {CodeXomics-ToolAgent-4B-v1},
  author = {Song, Lifu},
  year = {2026},
  license = {Apache-2.0},
  publisher = {Hugging Face},
  base_model = {Qwen/Qwen3.5-4B}
}

Available on FriendliAI

Dedicated Endpoints

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

Learn more

Model Details

Model Provider

CodeXomics

Model Tree

Base

Qwen/Qwen3.5-4B

Fine-tuned

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality