---
title: MemSifter-4B-Thinking
canonical_url: "https://www.modelscope.cn/models/zstanjj/MemSifter-4B-Thinking"
md_url: "https://www.modelscope.cn/models/zstanjj/MemSifter-4B-Thinking.md"
repository: zstanjj/MemSifter-4B-Thinking
last_updated: 2026-03-02
license: "MIT License"
pipeline_tag: text-generation
tasks:
  - text-generation
model_type:
  - qwen3
architectures:
  - Qwen3ForCausalLM
base_model:
  - Qwen/Qwen3-4B-Thinking-2507
base_model_relation: finetune
parameters: 4.4B
tensor_type:
  - BF16
library_name:
  - safetensors
  - pytorch
frameworks:
  - Pytorch
language:
  - en
  - zh
inference_backends:
  - "deploy_task text/emb"
  - "lmdeploy_turbomind 0.9.1"
  - "sglang 0.5.2"
  - "vllm 0.9.2"
downloads: 22
stars: 2
---

# MemSifter-4B-Thinking

> MemSifter-4B-Thinking - zstanjj 在 ModelScope 开源的模型。MemSifter-4B-Thinking

zstanjj/MemSifter-4B-Thinking 是 ModelScope 魔搭社区上的 4.4B 参数text-generation模型，采用 MIT License 许可，基于 Qwen/Qwen3-4B-Thinking-2507 构建，可用 deploy_task text/emb、lmdeploy_turbomind 0.9.1、sglang 0.5.2 部署。

- **Repository**: zstanjj/MemSifter-4B-Thinking
- **License**: MIT License
- **Tasks**: text-generation
- **Parameters**: 4.4B
- **Base model**: Qwen/Qwen3-4B-Thinking-2507
- **Inference backends**: deploy_task text/emb, lmdeploy_turbomind 0.9.1, sglang 0.5.2, vllm 0.9.2
- **Downloads**: 22
- **Stars**: 2
- **Last updated**: 2026-03-02

Source: https://www.modelscope.cn/models/zstanjj/MemSifter-4B-Thinking

---

# MemSifter-4B-Thinking

**MemSifter** is a lightweight generative session ranker trained with DAPO reinforcement learning. It serves as the core retrieval component of the [MemSifter](https://github.com/plageon/MemSifter) system—an LLM memory retrieval offloading framework based on outcome-driven proxy reasoning.

Given a user query and a set of candidate conversation sessions (pre-filtered by a dense embedding model), MemSifter-4B-Thinking performs fine-grained reranking to surface the sessions most relevant to the query, which are then passed as context to a downstream chat LLM.

## System Pipeline

```
Session Embedding  →  Session Ranking (MemSifter)  →  Chat LLM
   (bge-m3)           (generative reranker)          (any LLM)
```

1. **Session Embedding** — bge-m3 performs a coarse similarity pre-filter across all sessions.
2. **Session Ranking** — MemSifter (this model) performs fine-grained reranking of the pre-filtered candidates.
3. **Chat LLM** — the top-ranked sessions are assembled into a context window and passed to any OpenAI-compatible chat model.

## How to Use

Install the required packages:

```bash
pip install torch sentence-transformers vllm openai pyyaml loguru numpy pandas
```

Clone the repository and run the three-stage pipeline:

```bash
git clone https://github.com/plageon/MemSifter.git
cd MemSifter
```

```python
import json
from memsifter.toolkit import SessionEmbedder, SessionRanker, LLMChat

# Load one sample
with open("data/test_memory.json") as f:
    entry = json.load(f)[0]

question             = entry["question"]
haystack_sessions    = entry["haystack_sessions"]
haystack_dates       = entry["haystack_dates"]

# Initialise models
embedder = SessionEmbedder(model_path="models/bge-m3", device="cuda:0")
ranker   = SessionRanker(
    model_path="models/zstanjj/MemSifter-4B-Thinking",
    device="cuda:1",
)
chat = LLMChat(api_key="YOUR_KEY", base_url="YOUR_BASE_URL", model_name="YOUR_MODEL")

# Stage 1 — embedding pre-filter
top_sessions = embedder.get_top_sessions(
    question=question, sessions=haystack_sessions, dates=haystack_dates, top_k=20
)

# Stage 2 — generative reranking
ranked_sessions = ranker.rerank(
    question=question, pre_ranked_sessions=top_sessions, top_k=5
)

# Stage 3 — LLM answer
predicted_answer = chat.answer(question=question, ranked_sessions=ranked_sessions)
print(predicted_answer)
```

The `SessionRanker` uses [vLLM](https://github.com/vllm-project/vllm) for inference. It outputs a chain-of-thought inside `<think>...</think>` tags followed by a comma-separated session ranking inside `<ranking>...</ranking>` tags.

## Training

MemSifter-4B-Thinking is fine-tuned from **Qwen3-4B** using the **DAPO** reinforcement learning algorithm with a **task reward** formulation that combines:

- **Marginal Utility Reward** — rewards selecting sessions with diminishing redundancy.
- **Rank-Sensitive Reward** — rewards placing the most relevant sessions at higher positions.

Training data is bootstrapped from the MemSifter embedding pipeline on multiple conversational memory benchmarks (LoCoMo, LongMemEval, PersonaMem, etc.), using NDCG-based anchor sampling to construct RL training trajectories.

## Citation

```bibtex
@misc{memsifter2025,
  title     = {MemSifter: Offloading LLM Memory Retrieval via Outcome-Driven Proxy Reasoning},
  author    = {{Qwen Team}},
  month     = {February},
  year      = {2026},
  url       = {https://github.com/plageon/MemSifter}
}
```