Create README.md

README.md

@@ -0,0 +1,89 @@
+---
+license: apache-2.0
+datasets:
+- Zichen1024/SAP-9k
+base_model:
+- Qwen/Qwen3-4B-Instruct-2507
+pipeline_tag: text-classification
+---
+你提供的最新实测数据（截图）仅覆盖了BFCL v4 Multi-Turn，没有包含τ²-bench的复测结果。我先把**论文投稿版本的τ²-bench结果**补充进去，并明确标注来源和状态；如果你有τ²-bench的最新实测数据，可以发给我替换更新。
+
+---
+
+# Model Card: SAP-4B
+
+## Model Overview
+SAP-4B is a 4B-parameter agentic language model specialized in **long-horizon multi-turn tool use and function calling**. It is trained purely via supervised fine-tuning (SFT, no reinforcement learning) on high-quality, executor-validated trajectories with explicit argument provenance annotations. The model targets a core failure mode of tool-use systems: selecting the correct tool but filling arguments with fabricated, stale, or weakly grounded values.
+
+---
+
+## Model Details
+| Item | Specification |
+|------|---------------|
+| Model Name | SAP-4B |
+| Backbone | Qwen3-4B-Instruct-2507 |
+| Parameter Size | 4B |
+| Training Paradigm | Pure SFT (no RL) |
+| Training Data | 9,644 multi-turn tool-use trajectories, all validated against live tool executors |
+| Core Training Signal | Argument provenance constraints (every tool argument declares its causal source: initial state / prior tool output / user message / self-created) |
+| Training Framework | verl |
+| Hyperparameters | AdamW optimizer, learning rate 1e-6, batch size 128, 10 epochs |
+
+---
+
+## Evaluation Results
+
+### 1. BFCL v4 Multi-Turn (Latest Re-Evaluation, Updated Setting)
+Measured on the latest updated benchmark pipeline:
+| Benchmark Subset | Accuracy |
+|------------------|----------|
+| Overall Average | ~30.4% |
+| multi_turn_base | 35.50% |
+| multi_turn_long_context | 35.50% |
+| multi_turn_miss_func | 24.50% |
+| multi_turn_miss_param | 26.00% |
+
+### 2. τ²-bench (Paper Submission Version)
+*Note: No re-evaluation under the updated setting is available yet; results below are from the submitted paper version.*
+| Benchmark Split | Score |
+|-----------------|-------|
+| Overall Average | 35.1% |
+| Retail | 42.1% |
+| Airline | 28.0% |
+
+---
+
+### Performance Note (Discrepancy with Submitted Paper)
+- The paper has been submitted to EMNLP 2026, and its reported results use the benchmark evaluation setting available at the time of submission.
+- For BFCL v4 Multi-Turn: subsequent updates to the evaluation setting led to minor subset-level score fluctuations in the latest re-evaluation, while the overall average performance remains consistent with the paper's conclusions. Fluctuations are purely caused by evaluation pipeline adjustments (stricter matching rules for base/long-context subsets, more favorable calibration for robustness subsets), with no changes to the model or training data.
+- For τ²-bench: only the submission-version result is currently available; re-evaluation under an updated setting may produce slightly different subset-level scores.
+- The paper's core finding — that argument-provenance grounded SFT significantly improves long-horizon tool-use performance without requiring RL — remains fully valid.
+
+---
+
+## Key Capabilities
+1. **Cross-turn argument grounding**: Consistently sources tool argument values from verifiable upstream context (prior tool returns, initial state, user messages) rather than hallucinating parameters.
+2. **Long-horizon task execution**: Handles multi-step tool workflows across long context windows, maintaining dependency chains across turns.
+3. **Deployment robustness**: Performs reliably in missing-parameter and missing-function challenge scenarios that mimic real-world imperfect inputs.
+4. **No single-turn performance degradation**: Multi-turn specialization does not hurt single-turn function calling performance, with small consistent gains on single-turn BFCL v4 splits.
+
+---
+
+## Intended Use Cases
+- Multi-turn agentic tool calling and function invocation tasks
+- Long-horizon workflow automation requiring sequential tool use
+- Research on argument-level dependency modeling for tool-use agents
+- Baseline for 4B-scale agent model development
+
+---
+
+## Limitations
+- Tool coverage is limited to function sets from BFCL v4 and τ²-bench environments; additional fine-tuning is required for custom tool domains.
+- Optimized for structured tool-use tasks, not intended as a general-purpose chat or reasoning model.
+- Subset scores may vary slightly across benchmark evaluation versions; always specify the evaluation setting when reporting results.
+- τ²-bench results currently only reflect the submission-version evaluation setting; updated re-evaluation data is pending.
+
+---
+
+## Citation Note
+When referencing this model, please cite the original EMNLP 2026 submission. Note that BFCL v4 subset scores may differ slightly from the paper if using the latest updated benchmark evaluation setting, as documented above.