autmoate/cactus-needle2-reminder-dt

Needle 2 Reminder (German)

Task-specific fine-tune of Needle 2 for setting reminders and alarms in German. Handles absolute times ('Erinnere mich morgen um 8 Uhr an X'), relative offsets ('20 Minuten vorher'), and person-specific reminders. Extracts literal spans (target, when, relative, person).

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 ~589 German calendar examples (train split), evaluated on ~120 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": "reminder_parse",
    "description": "Extract the literal spans needed to set a reminder: either an absolute time, or a relative offset before a calendar event. Omit every field without evidence.",
    "parameters": {
      "type": "object",
      "properties": {
        "target": {
          "type": "string",
          "description": "What the reminder is about, exactly as written: an event title like 'Zahnarzt' or a subject like 'Wasser trinken'."
        },
        "when": {
          "type": "string",
          "description": "Absolute time phrase exactly as written, e.g. 'morgen um 8 Uhr', 'um 18:00'."
        },
        "relative": {
          "type": "string",
          "description": "Relative offset span exactly as written, e.g. 'drei Stunden vorher', '20 Minuten vor dem Termin'."
        },
        "person": {
          "type": "string",
          "description": "Person name exactly as written when the reminder is for someone else, e.g. 'Lisa'."
        }
      }
    }
  }
]],
    weights="autmoate/cactus-needle2-reminder-dt.cact",  # download from HF, pass local path
)

# Example: Erinnere mich 20 Minuten vor dem Zahnarzt
response = agent.complete("Erinnere mich 20 Minuten vor dem Zahnarzt")
print(response["function_calls"])

Evaluation Results (Base vs Fine-Tuned)

Metric	Base	Fine-Tuned
Full-frame Exact Match	0.0917	0.2333
Tool Call Accuracy	0.8679	0.7453
Field Precision	0.6742	0.9038
Field Recall	0.3886	0.4105
Field F1	0.4931	0.5646
Hallucinated Field Rate	36.4%	20.1%
False Positive Tool Rate	28.6%	28.6%
False Negative Tool Rate	13.2%	25.5%
Mean Latency (ms)	207	173

Training Data

• Training examples: 589 (deterministic, seedable)
• Evaluation examples: 120 (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).