greghavens/fable-code-next-step

Fabletron v2 — Arm A (verified)

Training corpus for greghavens/fabletron-v2-nemotron-3-super-120b. Rendered with the unsloth/NVIDIA-Nemotron-3-Super-120B-A12B chat template; every row is ready to tokenize as-is.

rows	11,381 (train 10,405 / validation 976)
trajectories	2,201
source	greghavens/fable-5-coding-and-debugging-traces @ 3dc1b76ae1e9a708f220d9ebc3301d844297cebb
license	cc-by-4.0
tokenizer	unsloth/NVIDIA-Nemotron-3-Super-120B-A12B

Length — this is what sets MAX_SEQ

p50	p90	p99	max
3,885	47,155	83,065	102,963

MAX_SEQ = 102,963 gives zero truncation. Anything lower silently drops the tail of the longest trajectories — and the tail is where the multi-step agentic behavior lives.

Supervision rule — read this before training on it

Each row is a cumulative prefix: the complete context through assistant turn k, where that final assistant message is the sole loss target. Earlier assistant turns are context, not targets.

target_start_token on every row marks where the loss span begins. Mask everything before it. Supervising every assistant span instead would both overweight early steps (turn 1 of an N-turn trajectory appears in all N rows) and leak, because later rows contain earlier rows' targets as context.

Fields

field	meaning
text	rendered trajectory, ready to tokenize
n_tokens	token count under unsloth/NVIDIA-Nemotron-3-Super-120B-A12B
target_start_token	first token of the loss span
split	train / val
trajectory_id	fable:<task> — never crosses splits
source_index	row index in the source snapshot
task, lang, category	provenance from the source corpus
assistant_step / assistant_steps	position k of N in the trajectory
target_tool_calls	tool calls in the target (parallel bundles stay native)

Split policy

Inherited from the source corpus and re-asserted at build time:

• Zero trajectory crossings. Every row of a task lands in one split — the builder raises if a task appears in both.
• Zero cross-split duplicate targets. No rendered text appears in both splits (0 exact-duplicate groups within splits are reported in BUILD_REPORT.json).

What this build removed from the source

The source redactor substitutes into raw JSON text rather than decoded strings, so its placeholders land inside code and prose — @pytest.fixture became \[REDACTED_EMAIL] (the pattern's local-part matched the n of a \n escape), and [REDACTED_SECRET] replaced Authorization: header literals in documentation and test fixtures.

rows dropped — placeholder inside the loss target	106 (84 train / 22 val, 41 tasks)
rows kept — placeholder in context only (masked, never trained)	2,546
tool-call arguments repaired (orphan \[ → [)	32

A placeholder in a context turn is masked and harmless. One inside the target would train the model to emit [REDACTED_SECRET] in code it writes, so those rows are dropped and the builder asserts that none survive. The escape repair is the only byte-level edit made to source content; it strips a backslash JSON does not define an escape for, restoring the literal the redactor intended.

Provenance and licensing

Derived solely from greghavens/fable-5-coding-and-debugging-traces (cc-by-4.0). No AGPL inheritance — this arm carries no copyleft obligation onward to models trained on it.

Reproducing this build

git clone https://github.com/greghavens/fabletron        # builder helpers
git clone https://github.com/greghavens/fabletron-v2 && cd fabletron-v2
python scripts/build_arm_a.py --revision 3dc1b76ae1e9a708f220d9ebc3301d844297cebb

BUILD_REPORT.json in this repo carries the full fingerprint: source commit, per-shard SHA-256, output SHA-256, row counts, and length distribution. scripts/check_fingerprint.py fails the training launch closed on any mismatch.