Rolodex-12B (12B active / 106B total MoE parameters, GLM-4.5-Air derivative) is
a search model: fine-tuned for agentic, tool-driven retrieval over personal and
team knowledge bases — multi-step search, memory lookup, and evidence
synthesis with tool use.
This model: LoRA fine-tune (r=32, α=64) with OAPL on agentic search
trajectories, merged into the base weights.
This repo contains the fully merged weights, plus the LoRA adapter under
adapter/. Because the base is the public GLM-4.5-Air, the adapter can also be
applied directly to the base model (see adapter/README.md, including the
router-bias step).
Rolodex-12B is also the stage-1 base of
PolyClerk-12B, our legal
work-product model.
Training
Method
OAPL, LoRA r=32 / α=64 (merged)
Framework
ms-swift (Megatron backend)
Epochs
1
Evaluation
On internal short-factoid retrieval QA over indexed knowledge bases, the
OAPL-trained model improves materially over the GLM-4.5-Air base while using
single-tree guided search (fewer tool calls than majority-vote baselines).
Internal numbers; independent benchmarks pending.
Usage
Requires ~200GB of weights (bf16). Serve with vLLM:
The chat template is included (chat_template.jinja). The model is trained for
tool-use loops (search/read tools) — it performs best inside an agentic harness
that exposes retrieval tools, rather than as a plain chat model.
Intended use & limitations
Intended for retrieval-augmented and memory-augmented agent workloads. It may
underperform general instruction models on open-ended chat. Verify retrieved
claims against sources; the model can still hallucinate under sparse retrieval.
