Scalably/qwen-agent-sft-experiments

Fine-tuning Qwen3.6-35B-A3B on production agent decisions: experiment log and results

The experiment log and results from our effort to make an open model a better worker for the agent platform we run for real businesses. The headline runs, X17 to X20, fine-tune Qwen/Qwen3.6-35B-A3B (MoE, 3B active) with LoRA in ms-swift on rented GPUs; each was preregistered with a pass bar before any GPU was rented, and the result rows apply that bar. Earlier rows (X01 to X16, R01) are reinforcement-learning and serving experiments on Qwen3.8-27B and Qwen3.5-9B through the Tinker training API, and S0 is a serving check; not all of those were preregistered. Failures are kept.

What is not here, on purpose: the training examples and the trained adapters. They are built from our clients' real agent sessions, and our clients own that data. This dataset publishes what we measured and what it cost.

• Write-up: Four LoRA runs on our own agent decisions (Hugging Face article)
• Built by: Scalably. One founder directs AI agents: Anthropic's Claude ran the pipeline, OpenAI's Codex and Claude reviewed every training record blind, and both families reviewed the results. No outside expert has reviewed this yet.

Subsets

Subset	Rows	What one row is
results (default)	72	one metric for one model arm in X17 to X20 (arms: base, LoRA seeds A/B, reasoning SFT sftr, dpo, x18, x19, x20); unit says rate or count
runs	45	one log entry (kind: preregistration, result, a stage name, or note), with its question, verdict and the cleaned full record
gpu_costs	4	GPU spend per training run

What moved

Run	Change	Before	After
X17	first SFT on reviewed production decisions: right tool and arguments	44.7%	64.4% (seed B; seed A similar)
X18	reasoning SFT (think, then act): right tool and arguments	36.7%	58.7% (DPO on top: 60.2%, not significant)
X18	loads the required skill before a gated tool	55.6%	98.4%
X19	same skill gate, on skills held out of training	54.9%	93.7% (X18 scored higher on this test; it proved easier than designed)
X19 → X20	correct first step after loading a held-out skill	47.4% (base)	65.8% in X19 (p 0.12 vs base, 0.0074 vs X18), lost again in X20 (36.8%)
X20	right tool and arguments / IFEval		63.6% / 0.775

What didn't work: DPO on top of reasoning SFT added nothing measurable (it trained on 158 of 301 pairs; longer pairs didn't fit in memory); imitation alone didn't make the model avoid calls that had failed in production; restoring everyday accuracy in X20 erased X19's procedure-following gain. Compare arms within one run only. Scorers and settings changed between runs, which is why the base numbers differ.

Metrics

Metric (prefix)	Meaning
ev1 / regular_ev1	right tool and arguments on held-out production decisions (regular steps)
heldout-tenant	same, on decisions from clients never seen in training
unseen_clients	same, on decisions from clients never seen in training
ev2_skill_required_loaded / seen_gate_L1	loads the required skill before using a gated tool (skills seen in training)
unseen_gate_L2 / gate_L2	same, on skills held out of training (their names still appear in system catalogs, and the block message names the skill)
unseen_gate_masked_L3 / masked_L3	same, with the skill name masked in the block message
pass4_*	the same gate passed on all 4 of 4 samples
clean_trajectory_joint	right tool and arguments on decisions from trajectories with no prior error
ev2_exact_repeat	with the error visible, repeats the exact failed call (lower is better)
numeric_string_args	count of numeric arguments sent as strings (lower is better)
malformed / no_tool_call	count of turns with no or unparseable tool call (lower is better)
first_step	correct first action right after loading a held-out skill
ifeval	IFEval prompt-level strict, 200 prompts, general instruction following
exact_failed_action_repeats	makes the call production made that then failed; the failure is not visible to the model (lower is better)
missing_calls	turns where a tool call was required but none was made (lower is better)

Cost

Four training runs cost about $327 in rented GPU time ($110.92 of it invoiced, the rest estimated from hourly price and box time), mostly 2x H200 boxes for one night each. Data review and teacher rationales ran on AI subscriptions at no marginal cost.

Limitations

• One seed per run except X17. Test sets are small (38 to 264 cases per line), so read the p-values in runs before trusting a gap.
• The evaluation sets come from one platform's traffic; they say little about other agent stacks.
• record_json is the original log with names, storage locations and approval notes removed. Some internal finding IDs remain and point to documents that aren't public.

Citation

@misc{scalably2026qwensft,
  title  = {Fine-tuning Qwen3.6-35B-A3B on production agent decisions: experiment log and results},
  author = {Lazi\'c, Pavle and {Scalably}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Scalably/qwen-agent-sft-experiments}
}

Contact: pavle@scalably.io. Built 2026-10-26.