qwen3-8b-tau-sft/README.md

---
license: apache-2.0
base_model: Qwen/Qwen3-8B
datasets:
  - fuvty/tau-bench-synthetic
language:
  - en
library_name: transformers
pipeline_tag: text-generation
tags:
  - tau-bench
  - tool-calling
  - agent
  - sft
  - cold-start
  - qwen3
---

# qwen3-8b-tau-sft

Non-thinking cold-start SFT of [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) for
[tau-bench](https://github.com/sierra-research/tau-bench) retail, trained on successful
synthetic retail trajectories from
[fuvty/tau-bench-synthetic](https://huggingface.co/datasets/fuvty/tau-bench-synthetic).

This is the **epoch-1** checkpoint, which is the one that passed the pre-RL gates and was
used to initialize GRPO. It is a *behavior-priming* checkpoint: it teaches the model to act
as a non-thinking, one-tool-call-per-turn tau-bench agent. It is not a finished tau-bench
policy on its own — see [Results](#results).

## What "non-thinking" means here

The tokenizer/chat template is a patched Qwen3 template whose assistant branch renders
`<|im_start|>assistant\n<think>\n\n</think>\n\n` + content verbatim, and whose generation
prompt ends with that same empty-think prefill. Because `</think>` is therefore *prompt-side*,
think-relapse is structurally impossible even at temperature 1.0. The patched
`chat_template.jinja` and the embedded copy in `tokenizer_config.json` ship with this repo,
so `apply_chat_template` reproduces the training format exactly.

Tool calls are emitted as inline Hermes text in the assistant content (the patched template
does **not** render structured `tool_calls` lists):

```
{optional user-visible text}
<tool_call>
{"name": "<fn>", "arguments": {<args object>}}
</tool_call>
```

At most one call per turn. Tool results are supplied as `role: "tool"` messages, which the
template wraps in `<tool_response>…</tool_response>` inside a user turn.

## Training data

795 reward-1.0 retail trajectories from `traj-GLM5` were filtered down to **595 rows over 98
tasks**:

- airline dropped entirely (it is a held-out zero-shot transfer eval)
- **12 of 180** synthetic retail tasks dropped by a contamination gate that compares
  write-action goal signatures against the shipped tau-bench retail task sets. All 12 were
  retail-**train** collisions; **zero** collided with retail-**test**
- 135 trajectories dropped for containing a turn with more than one tool call
- reasoning traces (`reasoning_content`) stripped; the tau2 greeting turn dropped

Loss falls on exactly `assistant content + <|im_end|>`. The assistant prefill header, the
inter-turn newline and every observation block are present but masked out, so the training
token stream is byte-identical to what an inference server is fed.

## Training

| | |
|---|---|
| base | Qwen/Qwen3-8B |
| framework | slime 0.3.0 (Megatron), 4×H100 |
| rows / batch | 595 / 32 |
| steps | 54 (18 per epoch × 3); **this repo is epoch 1 = 18 steps** |
| lr | 1e-5 cosine, min 1e-6, warmup 0.1 |
| parallelism | TP2, seq-parallel, bf16 |
| loss | token-level SFT on pre-tokenized rows |

Final loss 0.118 at epoch 3; this epoch-1 checkpoint was selected because it already passed
the boundary-token gate below and further epochs were not needed for priming.

## Gates

**Boundary-token probe** (does SFT actually train the turn-ending specials?):

| bucket | n | mean p | median | frac < 0.5 |
|---|---|---|---|---|
| `<\|im_end\|>` | 64 | **0.9625** | 0.9983 | 1.6% |
| `<tool_call>` / `</tool_call>` | 62 | 0.9786 | 0.9998 | 0.0% |
| content | 7108 | 0.8867 | 0.9994 | 10.0% |

**Format/geometry gate**: converted rows were replayed through the actual multi-turn rollout
code with the row's turns scripted as sampled outputs, asserting exact-token prefix alignment
of the generation prompt at every turn, loss-token equivalence, tool-parser round-trip, and
eos invariants — **595/595 passed**. Longest assistant turn: 613 tokens.

## Results

Evaluated with **greedy** decoding on tau-bench retail-test (115 tasks) and airline (50 tasks,
never trained on), with a `gpt-4.1-mini` user simulator.

| | retail-test | airline |
|---|---|---|
| **this checkpoint (SFT only)** | 0.043 | 0.020 |
| after 32 GRPO steps from this init | **0.322** | 0.060 |
| Qwen3-8B *thinking*, zero-shot (reference) | 0.391 | 0.240 |

Sampled (T=1.0, 4 samples/task) after those 32 GRPO steps, where the decoding pathology
is absent:

| | pass@1 | pass@2 | pass@4 |
|---|---|---|---|
| retail-test | 0.315 | 0.443 | 0.574 |
| airline (never trained on) | 0.140 | 0.237 | 0.360 |

**Read the SFT-only number carefully.** 0.043 greatly understates the policy: under sampling
(T=1.0) the same weights score **0.637** average success on retail-train with
**pass@16 ≈ 0.97**. The gap is a greedy-decoding pathology — 93% of greedy eval episodes
exhaust their 30-turn budget in a repetition loop rather than terminating. Subsequent GRPO
mostly *repaired that pathology* (greedy truncation 0.93 → 0.33, repetition 0.48 → 0.15)
rather than adding raw capability, which is why greedy retail-test climbs 7.5× while sampled
training reward barely moves.

So: this checkpoint is a strong **sampled** tool-use policy and a weak **greedy** one. Use
sampling, or RL from it.

## Intended use and limitations

Research artifact for tau-bench agent RL. Not a general assistant, not tuned for safety or
chat. Its system prompt at training time was the synthetic dataset's own
`<instructions>/<policy>` wrapper and its tool schemas were the dataset's tau2-style retail
variant, which differ in three tool names from tau-bench's own retail toolset
(`cancel_order` vs `cancel_pending_order`, `find_user_by_contact` vs `find_user_id_by_email`,
`find_user_by_name` vs `find_user_id_by_name_zip`). In practice the model conditions on
whatever schemas are given in context and transferred to the real tau-bench names with
essentially no penalty (5 parse failures across 165 eval episodes).

## License

Apache-2.0, inherited from both Qwen3-8B and the tau-bench-synthetic dataset.