MotifTech commited on
Commit
e6a92be
·
verified ·
1 Parent(s): d809c4d

Add BFCL multi-turn agent-tracer trajectories (2,310)

Browse files
Files changed (3) hide show
  1. .gitattributes +1 -0
  2. README.md +93 -0
  3. sft_all.jsonl +3 -0
.gitattributes CHANGED
@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
58
  # Video files - compressed
59
  *.mp4 filter=lfs diff=lfs merge=lfs -text
60
  *.webm filter=lfs diff=lfs merge=lfs -text
 
 
58
  # Video files - compressed
59
  *.mp4 filter=lfs diff=lfs merge=lfs -text
60
  *.webm filter=lfs diff=lfs merge=lfs -text
61
+ sft_all.jsonl filter=lfs diff=lfs merge=lfs -text
README.md ADDED
@@ -0,0 +1,93 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ task_categories:
4
+ - text-generation
5
+ language:
6
+ - en
7
+ tags:
8
+ - agent
9
+ - agentic
10
+ - tool-use
11
+ - function-calling
12
+ - multi-turn
13
+ - trajectories
14
+ - synthetic
15
+ - bfcl
16
+ size_categories:
17
+ - 1K<n<10K
18
+ configs:
19
+ - config_name: default
20
+ data_files:
21
+ - split: train
22
+ path: sft_all.jsonl
23
+ ---
24
+
25
+ # agent-tracer BFCL multi-turn trajectories
26
+ > ⚠️ **Work in progress.** This dataset and [AgentTracer](https://github.com/MotifTechnologies/AgentTracer) are under active development. The schema and contents may change without notice.
27
+
28
+ 2,310 multi-turn function-calling trajectories in OpenAI-style `conversations` format, with
29
+ `tools` schemas, per-step tool calls, and per-trajectory evaluation metadata. The record layout
30
+ follows [`Motif-Technologies/agenttracer`](https://huggingface.co/datasets/Motif-Technologies/agenttracer).
31
+
32
+ Each trajectory is a stateful, multi-turn task over the executable API classes of the
33
+ Berkeley Function Calling Leaderboard (BFCL v4, `bfcl-eval==2026.3.23`): file system, messaging,
34
+ posting, tickets, trading, travel, vehicle control and math. Tasks were newly written for this
35
+ dataset (no BFCL evaluation question, initial state or answer was used as a seed), every call
36
+ was executed against the real API classes, and only trajectories that passed the upstream BFCL
37
+ `multi_turn_checker` (plus stricter state checks on turns whose ground truth is empty) were kept.
38
+
39
+ | Variant | Rows | Content |
40
+ |---|---|---|
41
+ | `miss_func` | 1,259 | A function needed in one turn is withheld; the assistant must say it cannot do it without changing state, then finish once the function is added |
42
+ | `base` | 558 | Multi-turn requests completed with the provided functions |
43
+ | `long_context` | 493 | Same as `base`, executed with BFCL's long-context environment state |
44
+
45
+ All rows have `source.subset == "toolcall"`; the variant is not stored per row.
46
+
47
+ | Teacher model | Rows |
48
+ |---|---|
49
+ | `z-ai/glm-5.3-flash` | 2,244 |
50
+ | `moonshotai/Kimi-K3` | 66 |
51
+
52
+ ### Format notes
53
+
54
+ - No system prompt, no `submit` / `code_exec` tools: each user turn ends with an assistant reply
55
+ that makes no tool call, as in BFCL's FC evaluation. `extra.exit_status` is `completed`.
56
+ - Assistant turns carry the teacher's `thinking`. Turns where the model reported 0 reasoning
57
+ tokens have an empty `thinking`; trajectories with no reasoning at all were rejected.
58
+ - Tool descriptions carry BFCL's Python-syntax hint ("Note that the provided function is in Python 3 syntax.").
59
+ - Records of 32,768 tokens or more (Qwen3.5 chat template, reasoning kept on every turn) were removed.
60
+ - Tasks whose user turns share an 8-gram with any BFCL multi-turn evaluation question were removed.
61
+
62
+ ### Fields
63
+
64
+ - `conversations` — `user` / `assistant` / `tool` messages; assistant turns may carry `thinking` and `tool_calls`.
65
+ - `tools` — the tool schemas available for that trajectory.
66
+ - `source` — `{provider, name, split, subset, synthetic{synth, model, effort, ground_truth, system_prompt}}`.
67
+ - `extra` — `{task, sample_idx, attempt, steps, missing_tools, error, evaluation, exit_status}`.
68
+ - `category` — `agent` (2,310).
69
+
70
+ ## Licensing
71
+
72
+ **Dataset contribution — Apache-2.0.** The task texts, initial environment states and the
73
+ trajectory structure are released under the Apache License 2.0. They are procedurally generated
74
+ and fictional: names, accounts, files and messages are invented.
75
+
76
+ **BFCL environment and tool schemas — Apache-2.0.** The tool schemas and the executable API
77
+ classes come from `bfcl-eval` (Berkeley Function Calling Leaderboard,
78
+ [gorilla](https://github.com/ShishirPatil/gorilla/tree/main/berkeley-function-call-leaderboard)),
79
+ licensed under Apache-2.0. Keep that attribution if you redistribute this dataset.
80
+
81
+ **Generating models.** Trajectories are outputs of `z-ai/glm-5.3-flash` and `moonshotai/Kimi-K3`.
82
+ Check each model's license terms before training on or redistributing these outputs; for Kimi K3 see the
83
+ [Kimi K3 License](https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE).
84
+
85
+ ## Citation
86
+
87
+ ```bibtex
88
+ @misc{agent_tracer_bfcl,
89
+ title = {agent-tracer BFCL multi-turn trajectories},
90
+ year = {2026},
91
+ note = {Generated with z-ai/glm-5.3-flash and moonshotai/Kimi-K3 on BFCL v4 executable APIs}
92
+ }
93
+ ```
sft_all.jsonl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a30d53742188d14da4ba4669a7542bb52056fa4f7df40d5587f0f3104da7c907
3
+ size 70383164