Fabliq-8B-Agent-FromBase-Reasoning 🌊🔬🧠
Reasoning-expanded sibling of Fabliq-8B-Agent-FromBase. Adds general + deep reasoning (WithinUs + Helio) on top of the FromBase agentic foundation. Mirrors Fabliq-8B-Agent-Reasoning but for the FromBase ablation lineage.
🔬 Why this variant?
Two parallel lineages, both end with reasoning expansion:
| Lineage | Phase-1 | Phase-2 |
|---|---|---|
| ToolBench line | Fabliq-8B-Agent (ToolBench → Fable-5) |
Fabliq-8B-Agent-Reasoning |
| FromBase line (this) | Fabliq-8B-Agent-FromBase (base → Fable-5) |
Fabliq-8B-Agent-FromBase-Reasoning (this) |
Comparing the two Phase-2 endpoints answers: does ToolBench foundation matter for the final reasoning model, or is Fable-5 + reasoning enough?
🧪 Model details
| Architecture | Lfm2MoeForCausalLM (24 layers, 32 experts, 4 experts/token) |
| Parameters | ~8B total / ~1B active (MoE) |
| Context | 8,192 trained · 128K native |
| Precision | bfloat16 |
| Fine-tune type | Full-parameter SFT continuation |
| License | Apache 2.0 |
📚 Training data (Phase-2B only)
Same as Phase-2: WithinUs 135 + Helio 146 = 281 rows × 4 epoch.
🔧 Training procedure
| Hyperparameter | Value |
|---|---|
| Base | LLM-OS-Models/Fabliq-8B-Agent-FromBase |
| Schedule | 4 epochs, LR 3e-7 |
| Max sequence length | 8,192 |
| GPUs | 8× H200 (effective batch 64) |
| Train runtime | ~6 min |
🌳 Model tree
LiquidAI/LFM2.5-8B-A1B
├─ ToolBench line:
│ └─ LFM2.5-8B-A1B-Terminal-ToolBench-Full-SFT-1Epoch
│ └─ Fabliq-8B-Agent
│ └─ Fabliq-8B-Agent-Reasoning
└─ FromBase line (this):
└─ Fabliq-8B-Agent-FromBase
└─ Fabliq-8B-Agent-FromBase-Reasoning ← this model
📜 License
Apache 2.0.