| --- |
| license: apache-2.0 |
| --- |
| Model Name: Playbox-Time-Hammer |
|
|
| Architecture: Qwen3.5-0.8B-Base (initialized from Playbox_RFT_Checkpoint) |
|
|
| Parameters: ~0.8B |
|
|
| Developers: Epoch AI Labs |
|
|
| Intended Use: Domain-specific exact-arithmetic specialist exclusively for schedule and elapsed-time calculations. |
|
|
| Tool Integration: Trained to emit exact-arithmetic tool calls and consume results for the time_add(hhmm, [minutes...]) tool, bypassing unreliable internal time arithmetic. |
| |
| Training Pipeline: Multi-turn Tool-Integrated Reasoning (TIR) SFT using selective label masking (learning to emit tools and process tool responses, but ignoring prompts/tool injections during loss calculation). Max 80 updates. |
| |
| Performance: |
| Schedule Elapsed OOD Family: 100.0% accuracy (up from 0.0% with the direct RFT). |
| Limitations: Hyper-specialized. Suffers from complete catastrophic forgetting on general math distribution. Must only be invoked via a high-precision router for appropriate problems. |
