🔭 AetherSearch SFT
A compact search agent that learns to reason, retrieve, and answer
Fine-tuned from Qwen2.5-3B-Instruct on 2,600 complete agent trajectories, including direct-answer, single-search, and multi-search examples.
🏠 Project · 🧪 Training code · 📚 Dataset
🔌 Bring your own retriever. AetherSearch SFT is a search-agent policy, not a self-contained QA service. The host runtime must execute each
<search>...</search>request and return evidence inside<information>...</information>.
✨ Highlights
- 🔎 Search-native behavior — learns when and what to search before answering.
- 🔁 Direct-answer, single-search, and multi-search trajectories — trained on 600 direct-answer, 1,025 single-search, and 975 multi-search examples.
- 🧾 Evidence-in-the-loop reasoning — retrieved passages stay visible as context while being excluded from the training loss.
- ⚡ Compact 3B backbone — built on Qwen2.5-3B-Instruct for accessible experimentation and deployment.
- 🧪 Reproducible release — public trainer, launcher, data checksum, schema tests, and artifact manifest are included or linked.
🧠 How it works
Question
│
▼
<think>reason about what is missing</think>
│
▼
<search>focused retrieval query</search> ─────► Search / RAG backend
▲ │
└──── <information>retrieved evidence</information> ◄────┘
│
├── repeat the search loop when more evidence is needed
▼
<answer>evidence-grounded final answer</answer>
The model produces the reasoning, search, and answer spans. Your runtime owns
retrieval: parse a completed <search> span, run the query, append the result
as <information>, and resume generation until the model emits <answer>.
For direct answers, the model emits its final answer without a search or
<information> observation.
📊 Model at a glance
| Field | Value |
|---|---|
| 🧱 Base model | Qwen/Qwen2.5-3B-Instruct |
| 🧬 Base revision | aa8e72537993ba99e69dfaafa59ed015b17504d1 |
| 🏗️ Architecture | Qwen2 causal language model |
| 🔢 Parameters | 3,085,938,688 |
| 🎛️ Weight dtype | BF16 |
| 📏 Context | 32,768 positions; training sequences capped at 4,096 |
| 📚 Training data | 2,600 complete trajectories |
| 🔍 Trajectory mix | 600 direct-answer + 1,025 single-search + 975 multi-search trajectories |
| 🎓 Training stage | One full-trajectory SFT stage |
🚀 Quick start
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "muradil211/AetherSearch_SFT"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.bfloat16,
device_map="auto",
)
model.config.use_cache = True
model.eval()
💡 Loading the checkpoint is only the first step. For end-to-end use, wrap generation in the retrieval loop shown above and preserve the XML protocol exactly.
🧬 Checkpoint identity
This model was trained once on the 2,600 records in the canonical
final_sft_2600.jsonl dataset, using the same configuration as the public
AetherSearch SFT-2600 training code. The release contains the final model
artifacts and reproducible code, not server-local logs or optimizer state.
Dataset SHA-256
5619896ccc30bfb9d39c2676ec058cb59a0295c31082645153318102da0a7ec8
🧪 Training recipe
| Setting | Value | Setting | Value |
|---|---|---|---|
| Epochs | 1 | Learning rate | 2e-6 |
| Scheduler | Cosine | Global batch size | 24 |
| Precision | BF16 + TF32 | Max sequence length | 4,096 |
| Padding | Dynamic | Distributed training | DeepSpeed ZeRO-3 |
The training configuration matches the public SFT-2600 recipe: one epoch,
learning rate 2e-6, cosine scheduling, BF16, TF32, gradient checkpointing,
dynamic padding, effective global batch size 24, and DeepSpeed ZeRO-3. On the
three-worker training topology, per-device batch size 1 and gradient
accumulation 8 resolve to that global batch. The completed checkpoint is
exported as final_model/.
The public launcher is hardware-topology independent: it uses the devices made visible by the surrounding runtime and derives gradient accumulation to keep global batch 24 unchanged. It does not embed physical GPU IDs, node addresses, NCCL fabric settings, allocator tuning, or server-local paths.
🎯 Supervision contract
- ⬛ System, user, and question tokens are masked.
- ⬛ Complete
<information>...</information>spans are masked. - ✅ Assistant
<think>,<search>, and<answer>spans are supervised. - ✅ The final assistant
<|im_end|>token is supervised.
The trainer, launcher, configuration, checksum, and schema tests are published in the AetherSearch SFT directory.
📦 Files and integrity
The release contains two BF16 SafeTensors shards, the shard index, model and
generation configuration, tokenizer assets, this model card, the project logo,
and MODEL_MANIFEST.sha256. It intentionally excludes optimizer states,
intermediate checkpoints, training_args.bin, evaluation bundles, and all log
files.
After download, verify the release from its repository directory:
sha256sum -c MODEL_MANIFEST.sha256
⚠️ Limitations
Generated searches and answers can be incorrect, unsupported, or unsafe; retrieval and answer verification remain the caller's responsibility. No evaluation result is claimed by this model card.
📜 Terms
No additional blanket license is asserted here. Review the Qwen2.5-3B-Instruct license and the AetherSearch SFT data attribution and rights status before redistribution or downstream use.
Built for experiments in agentic search and retrieval-augmented reasoning. 🔎✨