| --- |
| license: apache-2.0 |
| --- |
| Playbox RFT Checkpoint (Generalist Base) |
|
|
| |
| Model Name: Playbox_RFT_Checkpoint |
|
|
| Base Architecture: Qwen3.5-0.8B-Base (24 transformer layers, mixed linear/full attention, hidden dimension 1024, grouped-query attention). |
|
|
| Parameters: ~0.8B |
|
|
| Developers: Epoch AI Labs |
|
|
| Intended Use: General mathematical reasoning, chain-of-thought generation, and pattern completion. Serves as the default, untouched path for standard mathematical queries (e.g., GSM8K) in the Routed Hammer ensemble. |
|
|
| Training Pipeline: |
| Stage 1 (Full-Parameter SFT): Supervised fine-tuning on GSM8K training data to establish reasoning chains and formatting. |
|
|
| Stage 2A (Periphery-Only GRPO): Group Relative Policy Optimization restricted to the final 8 transformer layers to preserve lower-layer capabilities while improving math output. Utilized strict math gating and frontier filtering. |
|
|
| Stage 2B (Identity GRPO Alignment): Final 6 layers trained on behavioral objectives (identity, anti-sycophancy, ambiguity handling) with no math training. |
|
|
|
|
| Performance: |
|
|
| GSM8K: 50.3% accuracy, 94.8% format compliance (outperforming comparators up to 9× its size like Qwen2.5-Math-7B-Instruct). |
|
|
| Adversarial OOD: 22.0% accuracy (without routing). |
|
|
|
|
|
|
| Limitations: Vulnerable to catastrophic forgetting. If fine-tuned further on tool-integrated reasoning (TIR), it completely loses general mathematical ability (dropping to 3.5% on out-of-distribution math). Behavioral alignment (identity) remains unsolved (0% preservation). |
