Your agent's failed traces are wasted fine-tuning data

I kept noticing the same thing while building agents: when the agent screws up —
picks the wrong tool, sends bad arguments, over-refuses something harmless — that
failure is genuinely useful. It's a concrete example of my model doing the wrong
thing, which is exactly what I'd want to fine-tune away.

And then it just... sits in the logs. Buried in LangSmith/Langfuse exports, mixed
with retries and PII and duplicates, annoying enough to clean up that I never did.

So I made a small tool for it: trace2train.
It's a local CLI (no server, no account) that turns failed agent traces into SFT/DPO
training data. Sharing it here in case it's useful to anyone else, and because I'd
like feedback.

The quick version

$ pip install trace2train
$ trace2train inspect --demo

19 traces → 16 failed → 38% dirty (PII/dupes/noise) → 14 trainable

failure types (trainable):
  wrong_tool        8
  bad_args          2
  over_refusal      2

inspect is just rules — no LLM, no API key. It tells you how much of your log
is actually usable before you pay for anything. Then convert turns the usable
failures into LLaMA-Factory-ready JSONL.

A concrete example

Here's a failure from the bundled demo. The agent has a get_weather tool but
reaches for calculator instead:

user:      What's the weather in Berlin?
assistant: [tool_call] calculator(expr="weather Berlin")
tool:      {"error": "invalid expression"}
assistant: I got 42.

wrong_tool — and the right fix is obvious from the trace (it should have called
get_weather). So convert can turn it into a training pair:

{
  "conversations": [
    {"from": "human", "value": "What's the weather in Berlin?"},
    {"from": "gpt",   "value": "[tool_call] get_weather(city=\"Berlin\")"}
  ]
}

Every record also carries provenance (source trace, original error, why it was
corrected) so the output isn't a black box.

The one design choice I actually care about

Most "turn your logs into training data" tools are eager — they'll happily
generate a confident answer for every failure so the row count looks big.

trace2train doesn't. It only writes a corrected sample when the right answer is
actually derivable from the trace. If the failure is "did the code pass its
tests?" or "is this fact true?" — things the trace can't confirm — it marks them
skipped instead of inventing something.

The reason is simple: a dataset full of confident-but-wrong "corrections" is worse
than no dataset. I'd rather hand you 7 pairs you can trust and 1 honest skip than
8 rows you can't.

That's the whole personality of the tool, really.

Other bits that came from actually using it

• PII redaction, dedup, and eval-set leak filtering are built in
• one LLM call per trace (cheap — DeepSeek by default, any OpenAI-compatible API works)
• --review to approve/reject each sample by hand before it's written
• --resume so a rate-limited re-run doesn't re-pay for calls you already made
• a health check that warns if your set is 90% one failure type (it'll just overfit)

Inputs today: LangSmith JSONL, a generic messages format, and Langfuse v4 snapshots.

Try it

pip install trace2train
trace2train inspect --demo               # 30s, no data or key needed
trace2train convert --demo --no-llm -o out
trace2train review -o out

Honest status

It's an early v0.1. I've validated it on a small slice of a public dataset and an
end-to-end Langfuse run on synthetic data — that's it. Not a benchmark, not an
accuracy claim. (Pretending otherwise would kind of defeat the whole "don't make
things up" premise.)

What I'd love feedback on: the failure-detection rules — what am I missing? — and
what trace sources you'd want supported next.

Repo: https://github.com/wane528/trace2train (MIT). If you try it on real
traces, I'd genuinely like to hear how it went.

Comments

[Mads Hansen]

The conservative skip behavior is the right instinct. One more guardrail I would add is causal labeling: a failed trace can come from the model, a transient tool/backend error, a stale tool catalog, or a policy decision that was correct at the time. Training all four as model mistakes can teach exactly the wrong lesson. Preserve the prompt, tool-catalog, policy and environment versions, and only emit a correction when a deterministic oracle or human reviewer validates it. For dataset splitting, group by canonical task/session before deduplication so retry variants cannot land on opposite sides of train and eval. And for DPO, keep the rejected path plus the validated correction; the contrast often carries more signal than a synthetic “perfect” SFT answer alone.