muradil211/ToolWeave-BFCL-Rollout-Case-Study

🧵 ToolWeave BFCL Formal-Training Rollout Case Study

This dataset publishes the complete raw on-policy rollout artifact from ToolWeave formal-training update 2, together with a focused real-rollout case study and deterministic K=16 peer-group analysis for runtime-interaction credit assignment. The records contain protocol failures and self-correction; they are raw reinforcement-learning trajectories, not curated demonstrations and not benchmark results.

Project: Muradil-mamat-211/ToolWeave

🧾 Formal-Training Update 2 Raw Dataset

data/raw_trajectories_update_2_512.jsonl is the complete raw trajectory artifact produced by ToolWeave formal training at update_2.

Field	Value
Training update	update_2
Records	512 trajectories
Prompt groups	32 groups × 16 rollouts
Format	JSONL, one trajectory per line
Role	Raw on-policy training artifact, not SFT data or a benchmark result

Each record preserves the runtime messages, parser/provenance metadata, reward records, questions, ground truth, and policy responses generated during the formal-training rollout.

🔎 Source identity

Field	Value
Original BFCL sample	multi_turn_base_156
Source JSONL line	10
Trajectory index	9
Global step / batch / epoch	2 / 1 / 0
Group/prompt UID	1b94ddc9-3612-48c4-acf2-7b755d72330f
Individual rollout ID	8516d0df-e6fb-4a67-969d-637bfd967e77
Rollout offset	9
Rollouts sharing the group UID	16
Source artifact SHA256	806b209cf7e02a1a20396fa833238fbe3bf9a2eacd794af7b3b8ed17ab6ba3e4

non_tensor.uid is the prompt/group identifier shared by the K=16 rollouts. It is not a unique trajectory ID. matchtir_provenance.rollout_id identifies the individual rollout.

🗂️ Complete stateful sample

The published JSON preserves the complete trajectory record: all five statefully connected BFCL user turns, runtime messages, interaction provenance, reward records, questions, ground truth, and policy responses.

User turn	Ground-truth calls
0	get_flight_cost, book_flight
1	retrieve_invoice
2	contact_customer_support
3	ticket_login, create_ticket
4	edit_ticket

The ground-truth call-count structure is [2, 1, 1, 2, 1].

🔁 Runtime-interaction recovery

ToolWeave's formal local temporal axis is the sequence of real non-answer runtime interactions within one BFCL user turn. One runtime interaction is one assistant generation followed by parser/environment handling. A valid final answer is excluded from the local sequence and receives global advantage only.

For User Turn 3, the selected rollout contains six non-answer runtime interactions:

j=0  parse_error  → P_j=[] → r_j=0
j=1  parse_error  → P_j=[] → r_j=0
j=2  parse_error  → P_j=[] → r_j=0
j=3  parse_error  → P_j=[] → r_j=0
j=4  parse_error  → P_j=[] → r_j=0
j=5  one valid tool-call action containing:
       ├── ticket_login
       └── create_ticket
     → call rewards [1.0, 1.0]
     → r_j=1.0

The parser-rejected generations remain real discount steps at immediate reward zero. Only the final successfully parsed calls enter whole-user-turn matching. The valid action contains one <tool_call> block with a JSON array of two calls, so it remains one temporal interaction and its call rewards are averaged.

📐 Frozen formal-credit semantics

The active Stage-3 mode is runtime_interaction_final:

• all successfully parsed calls in one BFCL user turn are concatenated with multiplicity preserved and matched once;
• call rewards are scattered back to their originating runtime interaction;
• an unparsed interaction has P_j=[] and r_j=0, but remains in the timeline;
• discounting uses real non-answer runtime depth j, with gamma=0.9;
• local peers share (group_uid, user_turn_id, runtime_interaction_index);
• ragged normalization uses unbiased sample standard deviation without zero-padding;
• singleton and zero-variance peer sets abstain with A_local=0;
• fusion is A_TW = A_RODS + A_local, with no averaging or post-fusion normalization.

For the special rollout:

immediate rewards:  [0, 0, 0, 0, 0, 1]
discounted returns: [0.59049, 0.65610, 0.72900, 0.81000, 0.90000, 1.00000]
peer support:       [16, 16, 1, 1, 1, 1]

Its fixed-denominator progress reward is R_P=0.8, and its global normalized advantage is A_RODS=-0.4966976345. Full-precision peer means, sample standard deviations, local advantages, and fused advantages are provided in the analysis files.

The complete K=16 User Turn 3 audit contains 36 rows—one for every real non-answer runtime interaction across the group. runtime_interaction_index and runtime_depth are the formal temporal fields. The older interaction_index, tool_attempt_index, r_t, and R_t fields remain only for backward compatibility; they do not define discounting or normalization.

The same deterministic audit verifies offset 2 with the runtime state checker. Its User Turn 3 calls match locally, while an earlier User Turn 0 omitted book_flight and used SAN instead of the ground-truth SFO for get_flight_cost. The resulting TravelAPI state mismatch makes the stateful User Turn 3 terminal score zero; this is neither a User Turn 3 parser failure nor a matching failure.

✅ Implementation and solver provenance

The values are reproduced by the current frozen ToolWeave Stage-3 formal-training implementation. The source trajectory remains unchanged; the production implementation replays its runtime/provenance records deterministically. The complete relevant pytest suite, K=16 regression, parser-error token-broadcast checks, and deterministic CPU trainer tensor-contract checks passed. No new formal training or checkpoint generation was performed for this documentation synchronization.

Solver provenance:

• the MatchTIR paper describes maximum-weight KM/Hungarian assignment;
• the audited MatchTIR public helper at commit 975c4535fbb86a49f21ff7d291a1fa822f827684 uses sorted positive non-conflicting edges;
• ToolWeave uses SciPy's true linear_sum_assignment(..., maximize=True) through its production matching module.

ToolWeave is an adaptation, not a literal MatchTIR implementation.

📦 Files

File	Contents
data/raw_trajectories_update_2_512.jsonl	Complete 512-trajectory raw artifact from formal-training update_2
data/multi_turn_base_156_rollout_offset_9.json	Complete original rollout, including all five BFCL user turns
analysis/user_turn3_k16_credit_summary.json	Full-precision K=16 matching, return, normalization, fusion, and state audit
analysis/user_turn3_k16_credit_summary.csv	Compact per-rollout summary
analysis/user_turn3_k16_full_interaction_advantage.json	One structured row per User Turn 3 non-answer runtime interaction
analysis/user_turn3_k16_full_interaction_advantage.csv	Exact tabular production-replay export

The raw JSONL is the complete update-2 source artifact; the smaller JSON/CSV files provide focused, reproducible analyses of the selected K=16 group.

🔐 Credential-like benchmark fixtures

Credential-like strings in the full record are synthetic BFCL benchmark fixtures, not production credentials. Questions, ground truth, and initial environment configuration were checked against the corresponding static BFCL source row. No Hugging Face token, GitHub token, SSH private key, cloud access key, or local server path is included.

🙏 Upstream context

The task data and execution environment derive from the public BFCL/EnvTuning infrastructure in AWorld-RL and the Berkeley Function-Calling Leaderboard. MatchTIR is referenced for the structural local-credit backbone; ToolWeave's runtime-depth and BFCL user-turn adaptations are project-specific.