xliaoyi/dbt-agent-sft

dbt agent SFT trajectories

Agent trajectories of Qwen3.5-35B-A3B (no thinking) and of the teachers gpt-5.6-luna and claude-opus-5.5 solving dbt / SQL data-engineering tasks, collected for self-distillation in the INF 385T LLM post-training course project. Layout follows allenai/tmax-sft: messages, tools, source, metadata.

Subset	Rows	Tasks
qwen3.5-35b-a3b_self_all	7,475	115
qwen3.5-35b-a3b_self_only_success	1,383	72

• Tasks: 115 train tasks (55 Spider2-DBT, 16 ADE-bench, 44 data-eng-bench analytics); the 20 held-out data-eng-bench test tasks of the v2 split are excluded (10 of the 44 data-eng-bench tasks here are test tasks in the later v3 split).
• Rollouts: 65 attempts per task, at most 100 turns, served with vLLM (FP8). Reward 1 = the task's tests pass.
• Tools: bash, read_file, write_file, edit_file (OpenAI function-calling format, JSON in tools).
• _only_success: rollouts with reward 1, the self-distillation SFT data. Training used only examples of at most 49,152 tokens under the Qwen3.5 chat template.
• metadata: benchmark, task, attempt, reward, success, stop reason, number of turns, wall-clock seconds, model.
• Twilio account SIDs appearing in tool output are replaced by AC_REDACTED_SID.

from datasets import load_dataset
ds = load_dataset("xliaoyi/dbt-agent-sft", "qwen3.5-35b-a3b_self_only_success", split="train")

Teacher trajectories

Rollouts of the two teacher models in the same harness, tools and format (provider-neutral chat messages; Claude's tool_use/tool_result blocks are converted). Hidden reasoning is not stored.

Subset	Rows	Test-split rows	Tasks
gpt-5.6-luna_all	760	90	135
gpt-5.6-luna_only_success	144	16	42
claude-opus-5.5_all	547	90	135
claude-opus-5.5_only_success	303	59	89

• Tasks: 71 training tasks (55 Spider2-DBT, 16 ADE-bench; at most 40 turns; Luna 8 and Opus 5 attempts each) and the 64 data-eng-bench analytics tasks (at most 100 turns, 3 attempts each).
• metadata.split: test for the 30 data-eng-bench tasks held out for evaluation in our v3 split, train otherwise. Filter out split == "test" before training if you evaluate on those tasks.
• metadata.run: Luna round1 (training tasks), eval; Opus train (training tasks), eval.
• metadata.reasoning: Luna ran with reasoning_effort=none; Opus with adaptive thinking at effort medium.

ds = load_dataset("xliaoyi/dbt-agent-sft", "claude-opus-5.5_only_success", split="train")
ds = ds.filter(lambda r: r["metadata"]["split"] == "train")

Qwen3.6-35B-A3B harness comparison (v2, after the environment fixes)

The same 4 configs rerun after fixing the environment: 134 validated tasks (105 train, 29 test; reference solution passes, empty solution fails; dbt-product-return-rate-analysis excluded), no internet for the agents (offline dbt package mirror), agents installed offline at pinned versions, anonymous workspaces, and a stricter semantic check next to each benchmark's own test.

Subset	Harness	Thinking	Rows	Test-split rows	Passed (reward)	Passed (strict)
qwen3.6-35b-a3b_qwen-think_v2	Qwen Code 0.26.5	on	536	116	128	95
qwen3.6-35b-a3b_qwen-nothink_v2	Qwen Code 0.26.5	off	536	116	125	100
qwen3.6-35b-a3b_alt-think_v2	altimate-code 0.12.8	on	536	116	139	113
qwen3.6-35b-a3b_alt-nothink_v2	altimate-code 0.12.8	off	536	116	106	91

• Same format and sampling as v1. metadata.reward is the benchmark's own test; metadata.strict is 1 only if that test and the strict check (outputs compared with the reference solution's) both pass.
• metadata.stop: completed, turn_limit, context_overflow, agent_timeout, api_error or crashed.

Qwen3.6-35B-A3B harness comparison (v1)

Rollouts of Qwen3.6-35B-A3B (vLLM, FP8) in two off-the-shelf agent harnesses run through Harbor 0.25.0, with thinking on and off: 135 tasks of the v3 split (105 train, 30 test) x 4 attempts per config.

Subset	Harness	Thinking	Rows	Test-split rows	Passed
qwen3.6-35b-a3b_qwen-think_v1	Qwen Code 0.26.5	on	540	120	126
qwen3.6-35b-a3b_qwen-nothink_v1	Qwen Code 0.26.5	off	540	120	102
qwen3.6-35b-a3b_alt-think_v1	altimate-code 0.12.8	on	540	120	142
qwen3.6-35b-a3b_alt-nothink_v1	altimate-code 0.12.8	off	540	120	98

• v1 = before the environment fixes. In this run the agents had internet access, two tasks had broken reference solutions (dbt-product-return-rate-analysis, dbt-fix-marketing-attribution), nine ADE-bench reference solutions pointed at a wrong path, and rewards are the benchmarks' own tests only. The v2 subsets above are the rerun with these fixed.
• Sampling: thinking on: temperature 1.0, top-p 0.95, top-k 20; off: temperature 0.7, top-p 0.8, top-k 20, presence penalty 1.5. At most 100 turns, 2-hour agent timeout; web tools removed.
• messages: Qwen Code rows come from Harbor's ATIF trajectory and keep the model's reasoning in reasoning_content; altimate-code rows are rebuilt from its JSON event log (text and tool calls; no reasoning text). The harness system prompts and tool schemas are not stored (tools is []); tool names are the harness's own (run_shell_command, bash, glob, ...).
• metadata: split (v3), reward, success, stop (Qwen Code: completed / turn_limit / api_error; altimate-code: its termination reason), exception (Harbor's label), harness, enable_thinking, num_turns, seconds.
• The full Harbor jobs (logs, verifier output) are on Harbor Hub; links in the GitHub README.

ds = load_dataset("xliaoyi/dbt-agent-sft", "qwen3.6-35b-a3b_alt-think_v1", split="train")

Code, teacher trajectories and evaluation results: https://github.com/xliaoyi/llms-post-train