CrowdMind/sft-mix-trace-v3

Freaky Labz SFT Mix v3

The token-balanced version. Same four corpora as sft-mix-trace-v2, but with the software-engineering data cut until the behaviour corpora are half the token budget instead of 0.8% of it.

13,695 rows, 70.0M tokens, 74 MB. One epoch of this is a general assistant that can also do software engineering, which is what v2 was not.

part	domain	rows	share of rows	tokens	share of tokens
conversations	not trace_*, not nemotron*	8,732	63.8%	6.20M	8.8%
agent traces	trace_<source>	2,767	20.2%	28.80M	41.1%
Nemotron agentless	nemotron_agentless	1,798	13.1%	14.29M	20.4%
Nemotron openhands_swe	nemotron_swe	398	2.9%	20.76M	29.6%
total		13,695		70.04M

Behaviour (conversations + traces) is 49.9% of the token budget. Nemotron is 50.0%. That is the whole point of this dataset and it is the one number to check before training.

How the balance was made, and what it cost

Every behaviour row is kept. The Nemotron side was cut to match: 2,196 of its 256,253 rows, 0.86% of upstream, for 35.0M of its 4,142M tokens.

Cutting Nemotron rather than repeating the behaviour rows is the whole decision. The other way round — repeating the behaviour half 118 times to lift it from 0.8% to 50% — would produce a 1.03M-row file of which 1.02M rows are duplicates, which is 118 epochs of the same 11,499 conversations inside a single run. The behaviour corpora took deliberate work to write; spending that instead of spending Nemotron rows is the better trade.

Which Nemotron rows were cut is decided by sorting on a hash of the row id and walking that order until the budget is spent. That is a uniform random sample: it preserves the length distribution, it keeps both Nemotron parts in proportion to their natural token mass (59.2/40.8, giving 398 trajectories and 1,798 single-turn answers), and it reproduces from the ids alone with no seed and no dependence on file order.

If you want more SWE data than this, go to v2 — it has all 256,253 Nemotron rows, 4.18 billion tokens, and the same behaviour corpora. v3 is not better, it is a different point on the trade.

Note the row split is nothing like the token split: conversations are 63.8% of the rows and 8.8% of the tokens, because a conversation is 622 tokens and a trajectory is 51,000. Balance is measured in tokens, which is what the model actually reads.

Length

approx tokens per row	rows	share
median	866
over 16k	1,002	7.3%
over 32k	485	3.5%
over 64k	84	0.6%
over 128k	3	0.0%

At --max_length 131072 three rows of 13,695 are dropped. The median row is 866 tokens, so this is an ordinary SFT shape with a small long tail: fine for 16k if you want throughput (7.3% dropped), fine for 131072 if you do not.

Format

Same eleven columns as every other dataset here, and the same three properties worth restating:

• Tool calls are rendered into the assistant turn in Qwen's native XML, with the structured form in tool_calls_json. Qwen3.5's template iterates tool_call.arguments|items, so the OpenAI-convention JSON string raises TypeError: Can only get item pairs from a mapping; datasets cannot store a real mapping either, because pyarrow's map type is not loadable by datasets; and this template has no tool_call role branch. 63,527 tool calls across 3,571 rows, every one of them verified byte-for-byte against what the template produces for the equivalent structured form.
• Three system prompts. Conversations and traces carry the Fred persona; openhands_swe carries the OpenHands agent prompt, because every trajectory was produced under it; agentless had none of its own and gets the Fred persona, which is what its rows describe anyway. harness says which is which.
• images is empty on every row. difficulty is 0 outside the conversations, meaning "not rated".

Splits

split	rows
train	13,421
validation	274

1-in-50, assigned by a stable hash of the row id so it does not depend on read order. The same rule the Nemotron parents use, and it lands on 274 rows because the behaviour corpora contribute almost all of them.

Validation

Structural checks over all 13,695 rows: system first, assistant last, uniform message keys, no residual tool_calls field, every call covered by its own row's tools column, tool_calls_json indices in range and pointing at assistant turns, rendered marker count equal to call count, no duplicate ids.

The chat-template pass is not sampled at this size. All 13,695 rows render through the real Qwen/Qwen3.5-0.8B tokenizer: 0 failures, and all 3,571 tool rows match the template's own structured rendering exactly. Two rows have more tool-call markers in their text than they have calls because the source quotes tool-call markup; they are reported, not hidden.

Loading it

from datasets import load_dataset

ds = load_dataset("CrowdMind/sft-mix-trace-v3")

behaviour = ds["train"].filter(lambda r: not r["domain"].startswith("nemotron"))
swe        = ds["train"].filter(lambda r: r["domain"].startswith("nemotron"))

74 MB, and it loads in seconds, which is the other thing worth saying about this one.

Training it

CUDA_VISIBLE_DEVICES=0 swift sft \
  --model Qwen/Qwen3.5-4B \
  --dataset CrowdMind/sft-mix-trace-v3 \
  --train_type lora \
  --max_length 131072 \
  --loss_scale default+ignore_empty_think \
  --num_train_epochs 3 \
  --per_device_train_batch_size 4 \
  --gradient_accumulation_steps 4 \
  --learning_rate 1e-4 \
  --lora_rank 16 \
  --lora_alpha 32 \
  --target_modules all-linear \
  --output_dir output

• Three epochs, not one. At 70M tokens this is a small dataset, and one epoch leaves most of it unseen by the optimiser.
• --max_length 32768 drops 3.5% of rows and is a reasonable speed/faithfulness trade here; 131072 keeps all but three.
• --loss_scale ignore_empty_think matters for the conversations, the traces and agentless. It does nothing for openhands_swe, which has no think blocks — its reasoning arrives as calls to a think tool, and there are only 398 of those rows in this mix.
• The behaviour corpora are 8.8% of tokens by row count but 49.9% by tokens, so a single epoch already sees them properly. No upsampling flag is needed.

Provenance and licence

Same three provenances as v2, at 0.86% of the Nemotron rows:

• Conversations are synthetic and fully invented, on the reserved .test domain.
• Traces are other people's published agent sessions, anonymised (8,292 ids, 4,410 email addresses, 1,528 ip addresses and 46,950 paths rewritten) with 22 credential-bearing sessions removed. Check each source's licence.
• Nemotron is NVIDIA's published dataset under cc-by-4.0, redistributed as published, not anonymised.

other for the mix as a whole. The Fred persona is in three of the four parts; strip messages[0] and set --system if you would rather not train it.