deepaksamuel-cuk/trkfit-llm-datasets

trkfit-llm datasets

Datasets for the track-fitting language models, such as deepaksamuel-cuk/trackfit-llm-small and deepaksamuel-cuk/trackfit-llm-large. Track fitting is treated as translation: a list of encoded detector hits is translated into a call that regenerates them.

config	events	tracks per event	used for
test-extended	20,000	1	trackfit-llm-small
test-mixedtrack-extended	20,000	1, 2 or 3	trackfit-llm-large

test-extended: synthetic single-track test set (20,000 events)

This is the test set used to evaluate the extended small model (trackfit-llm-small, epoch 1740). The results are in that model card.

Files (same events, same order):

• test-extended/test-extended.jsonl: one JSON object per line, with fields input_hits, target_call, n_tracks and class.
• test-extended/test-extended.csv: columns event, class, n_tracks, input_hits, target_call. input_hits is a JSON list stored as a string.

{"input_hits": [7, 40, 74, 107, 141, 175, 208, 242, 275, 309, 343, 376], "target_call": "gen_evt(slp=[1.58], icpt=[6.75], add=[], rem=[])", "n_tracks": 1, "class": "clean"}

Detector and encoding

• Geometry: 12 layers × 32 strips.
• Hit encoding: each hit is one integer, layer * 32 + strip, in [0, 383].
• Input hits: input_hits is sorted and deduplicated, with 4–17 hits per event.
• Target call: target_call is the ground-truth call:

gen_evt(slp=[<slope>], icpt=[<intercept>], add=[<noise hit codes>], rem=[<missing hit codes>])

• Clean track: the track hits strip round(slp * layer + icpt) in every layer it crosses. slp is in strips/layer, and icpt is in strips at layer 0.
• Noise: add lists the noise hits added to the input, and rem lists the track hits removed from it.

How it was built

• Classes: 5,000 events each of clean, add, rem and addrem. class is the class the event was generated as.
• Noise lengths: add length is drawn from 2–5 and rem length from 5–8. 85% of the noise hits are adjacent to the track, and at least 5 hits remain after removal.
• Track parameters: (slp, icpt) pairs are drawn uniformly from the extended track list, extended-allowed.csv. That list has 41,497 pairs, one per distinct clean hit pattern: slopes [-8.14, 8.17] strips/layer and intercepts [-64.92, 95.92] strips.
• Independence: no input repeats within the set, and none is shared with the model's validation or earlier test sets. Training events are generated fresh every epoch from the same track list.

Loading

from datasets import load_dataset

ds = load_dataset("deepaksamuel-cuk/trkfit-llm-datasets", "test-extended", split="test")
print(ds[0])

test-mixedtrack-extended: synthetic 1-, 2- and 3-track test set (20,000 events)

This is the test set used to evaluate the large extended model (trackfit-llm-large, epoch 320). The results are in that model card. Detector, hit encoding and the meaning of slp, icpt, add and rem are the same as for test-extended above; the call simply lists one slope and one intercept per track.

Files (same events, same order):

• test-mixedtrack-extended/test-mixedtrack-extended.jsonl: one JSON object per line, with fields input_hits, target_call and n_tracks.
• test-mixedtrack-extended/test-mixedtrack-extended.csv: columns event, n_tracks, input_hits, target_call. input_hits is a JSON list stored as a string.

{"input_hits": [10, 41, 56, 105, 119, 136, 138, 151, 184, 216, 232, 233, 263, 294, 295, 313, 314, 315, 345, 377], "target_call": "gen_evt(slp=[-0.3, 0.22], icpt=[9.89, 22.57], add=[41, 56, 136, 138, 233, 294, 314, 315], rem=[23, 42, 55, 73, 87, 137, 168, 200, 248, 280, 327, 359])", "n_tracks": 2}

Contents

tracks	events	hits per event	max noise hits (add)	max missing hits (rem)
1	6,666	4–16	5	7
2	6,684	9–32	10	14
3	6,650	13–44	15	21

input_hits is sorted and deduplicated. Tracks in a call are ordered by (intercept, slope).

How it was built

• Track parameters: every track's (slp, icpt) is drawn uniformly from the model's track list, allowed.csv, 41,497 pairs: assets/extended-allowed.csv without the pair (-0.04, -0.76), whose line never enters the detector. Slopes in this set span [-8.14, 8.14] strips/layer and intercepts [-63.67, 95.68] strips.
• Noise per track: each track gets a class, clean : add : rem : addrem = 1 : 1 : 1 : 8. Noise hits: 1–4 for add, 2–5 for addrem, about 85% of them in the same layer and ±1 strip from that track. Missing hits: 3–7 for rem, 5–8 for addrem, always leaving at least 5 hits on the track (so at most 7 in practice). In 2- and 3-track events the per-track add and rem hits are merged into one add and one rem list.
• Separable tracks: in 2- and 3-track events every track keeps at least 6 clean hits that no other track shares; missing hits are removed only from those, and noise hits never fall on any track's clean hits.
• Independence: no input repeats within the set, and none is shared with the model's 2,000-event validation set. Training events are generated fresh every epoch with the same recipe.

Loading

from datasets import load_dataset

ds = load_dataset("deepaksamuel-cuk/trkfit-llm-datasets", "test-mixedtrack-extended", split="test")
print(ds[0])
print(ds.filter(lambda e: e["n_tracks"] == 3)[0])