autmoate/cactus-needle2-calendar-read-dt

Needle 2 Calendar Read (German)

Task-specific fine-tune of Needle 2 for reading and querying calendars in German. Handles questions like 'Was steht morgen an?', 'Wann hat Lisa Termine?', 'Wann sind Lisa und Max frei?'. Extracts literal spans (query_span, when, person, persons, target) from natural language queries.

Base Model

This model is a fine-tune of Cactus-Compute/needle2, a 45M-parameter tool-calling model that runs in 28MB of RAM. See the original model card for details on the Simple Attention Network architecture, deployment targets, and benchmarks.

What Was Done

LoRA fine-tuning (rank 16, alpha 32) on ~1157 German calendar examples (train split), evaluated on ~240 held-out German examples (eval split).

• Backend: JAX 0.10.2 on CUDA (Jetson AGX Orin 64GB, sm_87)
• Training: 10 epochs, lr 1e-4, batch size 16, cosine decay with warmup
• Numerics: Quantization-aware training (CQ mixed STE + A8)
• Export: W4A8 (4-bit weights, 8-bit activations)
• Size: ~23 MB per .cact archive

Usage

import needle

# Load the fine-tuned calendar model
agent = needle.Needle(
    tools=[[
  {
    "name": "calendar_read",
    "description": "Answer a calendar question: list upcoming events, look up one event, show a person's calendar, or find joint free time slots. Extract only literal spans that appear in the query. Omit every field without evidence.",
    "parameters": {
      "type": "object",
      "properties": {
        "query_span": {
          "type": "string",
          "description": "The question-intent span copied verbatim from the query, e.g. 'Was steht an', 'frei', 'Zeig', 'Welche Termine'."
        },
        "when": {
          "type": "string",
          "description": "Time scope phrase exactly as written, e.g. 'morgen', 'diese Woche', 'heute Nachmittag'."
        },
        "person": {
          "type": "string",
          "description": "Person name exactly as written when the question is about one person, e.g. 'Lisa'."
        },
        "persons": {
          "type": "string",
          "description": "Person names exactly as written when the question is about several people, e.g. 'Lisa und Max'."
        },
        "target": {
          "type": "string",
          "description": "Event title to look up, exactly as written, e.g. 'Zahnarzt' in 'Wann ist Zahnarzt?'."
        }
      }
    }
  }
]],
    weights="autmoate/cactus-needle2-calendar-read-dt.cact",  # download from HF, pass local path
)

# Example: Was steht morgen an
response = agent.complete("Was steht morgen an")
print(response["function_calls"])

Evaluation Results (Base vs Fine-Tuned)

Metric	Base	Fine-Tuned
Full-frame Exact Match	0.1167	0.1750
Tool Call Accuracy	0.8821	0.9575
Field Precision	0.4417	0.8413
Field Recall	0.1165	0.3495
Field F1	0.1843	0.4938
Hallucinated Field Rate	8.4%	5.8%
False Positive Tool Rate	14.3%	21.4%
False Negative Tool Rate	11.8%	4.2%
Mean Latency (ms)	177	145

Training Data

• Training examples: 1157 (deterministic, seedable)
• Evaluation examples: 240 (held-out values and phrasings)
• Language: German only
• Negative examples: ~12% (cross-task + off-topic)
• Grounding: Every argument value is a literal substring of the query

Datasets generated with a template-based dataset builder (needle-only/calendar_ft/build_dataset.py).

Citation

If you use this model, please cite both the base model and this fine-tune:

@misc{needle2_2026,
  title        = {Needle 2: A 45M-Parameter Foundation Tool-Calling Model for Tiny Devices},
  author       = {Ndubuaku, Henry and Mosoyan, Karen and Mroz, Jakub and Cylich, Noah and
                  Kumar, Satyajit and Sandhu, Parkirat and Shemet, Roman and Lee, Justin H.},
  year         = {2026},
  organization = {Cactus Compute, Inc.},
  howpublished = {\url{https://github.com/cactus-compute/needle}}
}

@misc{autmoate_calendar_ft_2026,
  title        = {Task-Specific Needle 2 Fine-Tunes for German Calendar Operations},
  author       = {autmoate},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/autmoate}},
  note         = {LoRA fine-tune of Cactus-Compute/needle2 on German calendar data}
}

License

Apache 2.0 (inherited from the base model).