---
license: apache-2.0
language:
- en
library_name: transformers
pipeline_tag: text-generation
base_model: JaspinXu/KnowWhenToHandOff-Qwen3-4B-SFT
base_model_relation: finetune
datasets:
- JaspinXu/KnowWhenToHandOff-data
tags:
- agents
- tool-use
- customer-service
- reinforcement-learning
- grpo
- tau2-bench
---

# KnowWhenToHandOff-Qwen3-4B-R2 (GRPO, Responsible-AI reward)

> **Research artefact — not for deployment.** Trained and evaluated only with
> simulated users on a public benchmark. It has not been tested with real
> customers, real data or real business systems.

**Code** [github.com/JaspinXu/KnowWhenToHandOff](https://github.com/JaspinXu/KnowWhenToHandOff) · **Data** [KnowWhenToHandOff-data](https://huggingface.co/datasets/JaspinXu/KnowWhenToHandOff-data) · **Family** [SFT](https://huggingface.co/JaspinXu/KnowWhenToHandOff-Qwen3-4B-SFT) · [R1](https://huggingface.co/JaspinXu/KnowWhenToHandOff-Qwen3-4B-R1) · [R2](https://huggingface.co/JaspinXu/KnowWhenToHandOff-Qwen3-4B-R2)

**R2** is Qwen3-4B-Instruct-2507 trained with SFT and then multi-turn GRPO on the τ²-bench airline and retail domains, with a reward that adds penalties for rule violations, unnecessary hand-offs and missed hand-offs. It has the best hand-off F1 of the study (0.812) with 10 % over-escalation, at the same task success as its SFT starting point.

## Model family

| Model | Training | Hand-off F1 | Over-escalation | Use it for |
| --- | --- | --- | --- | --- |
| [SFT (B2)](https://huggingface.co/JaspinXu/KnowWhenToHandOff-Qwen3-4B-SFT) | LoRA SFT on 790 filtered teacher dialogues | 0.768 | 0.016 | starting point; precise but misses hand-offs |
| [R1](https://huggingface.co/JaspinXu/KnowWhenToHandOff-Qwen3-4B-R1) | GRPO from B2, task reward only | 0.802 | 0.462 | studying the outcome-reward loophole |
| [R2](https://huggingface.co/JaspinXu/KnowWhenToHandOff-Qwen3-4B-R2) | GRPO from B2, task reward + violation and hand-off penalties | **0.812** | 0.103 | the best-calibrated hand-off model |

Task success on the official test tasks is the same for all three within
noise (0.33–0.35).

## Quick start

A standard Qwen3 chat model with tool calling. It was trained with the
τ²-bench domain policy as the system prompt and the domain's tools, including
`transfer_to_human_agents`; for faithful behaviour run it inside the evaluation
harness of the [code repository](https://github.com/JaspinXu/KnowWhenToHandOff).

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "JaspinXu/KnowWhenToHandOff-Qwen3-4B-R2"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
    repo, torch_dtype="auto", device_map="auto"
)

transfer = {
    "type": "function",
    "function": {
        "name": "transfer_to_human_agents",
        "description": "Transfer the user to a human agent.",
        "parameters": {
            "type": "object",
            "properties": {"summary": {"type": "string"}},
            "required": ["summary"],
        },
    },
}
messages = [
    {"role": "system", "content": "<airline or retail policy>"},
    {"role": "user", "content": "I need the unaccompanied-minor service."},
]
inputs = tokenizer.apply_chat_template(
    messages, tools=[transfer], add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)
output = model.generate(inputs, max_new_tokens=512)
print(tokenizer.decode(output[0][inputs.shape[-1]:], skip_special_tokens=True))
```

Serve it with vLLM (OpenAI-compatible API, tool calls parsed):

```bash
vllm serve JaspinXu/KnowWhenToHandOff-Qwen3-4B-R2 \
  --enable-auto-tool-choice --tool-call-parser hermes
```

## Evaluation

Test split, protocol frozen in [ADR-005](https://github.com/JaspinXu/KnowWhenToHandOff/blob/main/docs/adr/ADR-005-evaluation-protocol.md); 95 % bootstrap intervals over tasks.

| Metric | This model | B2 (SFT start) | Source |
| --- | --- | --- | --- |
| Success (pass^1), test_main | 0.342 [0.245, 0.449] | 0.327 [0.224, 0.434] | [main table](https://github.com/JaspinXu/KnowWhenToHandOff/blob/main/reports/main/main_table.md) |
| Paired Δ success vs B2 | +0.015 [−0.056, 0.082] | — | [paired differences](https://github.com/JaspinXu/KnowWhenToHandOff/blob/main/reports/main/paired_differences.md) |
| pass^4 | 0.122 [0.041, 0.224] | 0.102 [0.041, 0.204] | [main table](https://github.com/JaspinXu/KnowWhenToHandOff/blob/main/reports/main/main_table.md) |
| Violation rate (incl. off-reference writes) | 0.444 | 0.469 | same |
| Hand-off F1 / precision / recall, test_handoff | **0.812** / 0.886 / 0.750 | 0.768 / 0.976 / 0.633 | [`R2-final-test_handoff-s0-20261007`](https://huggingface.co/datasets/JaspinXu/KnowWhenToHandOff-data/blob/main/runs/R2-final-test_handoff-s0-20261007/metrics.json) |
| Over-escalation, test_handoff | 0.103 | 0.016 | same |
| Worst-style success / style gap | 0.316 / 0.112 | 0.327 / 0.112 | [styles](https://github.com/JaspinXu/KnowWhenToHandOff/blob/main/reports/main/styles.md) |

**Known behaviour.** R2 removes R1's over-escalation but sometimes makes the opposite error on out-of-scope requests: it refuses and tells the user to contact the company directly without calling the transfer tool, or substitutes an unrelated write (under-escalation 0.25). Once it justified a hand-off with a policy that does not exist.

## Training details

- Base model: Qwen/Qwen3-4B-Instruct-2507 (revision
  `cdbee75f17c01a7cc42f958dc650907174af0554`), Apache-2.0.
- Training: GRPO from the SFT checkpoint B2 with the composite reward ([`configs/reward/r2.yaml`](https://github.com/JaspinXu/KnowWhenToHandOff/blob/main/configs/reward/r2.yaml): task reward − 0.25 × weighted violations − 0.5 × unneeded transfer − 1.0 × missed transfer − 0.1 × format errors); LoRA rank 64, alpha 128, lr 1e-5; 8 tasks × 8 rollouts per step, 60 steps, no KL term.
- Training run `R2-s1-20261006` (seed 1); exported step-60 checkpoint
  `sha256:be6198a0c51773f52ba4f20c9605a5931f23b0206e92a9fe0521a93761348c96` (merged bf16, Hugging Face format).
- Evaluation code at git `462b5d17b4e9add637a2ee7802974c720b625a5b`; τ²-bench
  `fc0055dc4e0a316c3f83133267fbd6faaa770992`; all sources in
  [reports/main/sources.json](https://github.com/JaspinXu/KnowWhenToHandOff/blob/main/reports/main/sources.json).

## Intended use

Studying when a tool-using customer-service agent should hand off to a human,
in the τ²-bench airline and retail domains. Out of scope: any production or
customer-facing use, other domains, and decisions about real people.

## Training data

Simulated conversations on τ²-bench v1.0.1 train-split tasks (a dev split was
carved out for selection) and hand-off variants derived from them (explicit
requests for a human, out-of-scope requests, hard negatives); the user
simulator is Qwen3-30B-A3B-Instruct-2507. No real user data. Test tasks were
never used for training or selection.

## Limitations

- One user simulator (Qwen3-30B-A3B-Instruct-2507, which also generated the
  SFT data) and one training seed; results may not transfer to real users or
  other simulators, and the intervals cover task sampling, not training variance.
- Small test split (49 official tasks; retail uses the 74 tasks whose reward
  needs no LLM judge, so retail numbers are not comparable with leaderboards).
- Communication styles describe how users write, not who they are; no claim is
  made about any demographic group.
- Policy compliance is checked by five deterministic rules that cover only
  part of each domain policy; citation rate is a keyword rule (it counted one
  invented policy as a citation).
- Observed reward-hacking behaviours:
  [docs/rl-audit.md](https://github.com/JaspinXu/KnowWhenToHandOff/blob/main/docs/rl-audit.md).

## Licences

The weights are a LoRA fine-tune of Qwen/Qwen3-4B-Instruct-2507 and inherit
its Apache-2.0 licence. τ²-bench code and data (pinned commit
`fc0055dc4e0a316c3f83133267fbd6faaa770992`) are used under the licence in
its `LICENSE` file;
released dialogue samples are τ²-bench-derived synthetic conversations with
no real user data.

## Citation

Model DOI: [10.57967/hf/10823](https://doi.org/10.57967/hf/10823). Cite the model and the project:

```bibtex
@misc{knowwhentohandoff2026r2,
  title     = {KnowWhenToHandOff-Qwen3-4B-R2},
  author    = {{KnowWhenToHandOff contributors}},
  year      = {2026},
  publisher = {Hugging Face},
  doi       = {10.57967/hf/10823},
  url       = {https://huggingface.co/JaspinXu/KnowWhenToHandOff-Qwen3-4B-R2}
}

@misc{knowwhentohandoff2026,
  title  = {Know When to Hand Off: Multi-turn Reinforcement Learning for
            Trustworthy Customer-Service Agents},
  author = {{KnowWhenToHandOff contributors}},
  year   = {2026},
  howpublished = {\url{https://github.com/JaspinXu/KnowWhenToHandOff}}
}
```