jscpdv5

karpathy/nanochat

The best ChatGPT that $100 can buy.

Python★ 60,004+321 that day#7 on trending

Analyzed November 3, 2026 at 92d63d4 (master) with jscpd v5 in 1.1s.

D42.6/100
Project health8,159 lines of code (S)
Duplication84
1.7% in python, bash
Dead code12
28%
Complexity72
24.6%

The score is jscpd's own --health model: markup, data and text left out, tests kept. The numbers below are the trending scan's.

1.18%duplicated code · 84 of 7,099 lines
8clones in code
33code files of 49 scanned
1.84%duplicated over all files · 22 clones

Most complex files

Dead code 28% unused · 11 findings

The code number leaves out tests, fixtures, docs, examples, data files, styles, translations and vendored code.

Largest clones in the code

  • 8 lines in 2 placespython86 tokens
    scripts/chat_eval.py:67–74scripts/chat_eval.py:141–148
    67# Aggregate results across all ranks
68if ddp:
69    num_passed_tensor = torch.tensor([num_passed], dtype=torch.long, device=device)
70    total_tensor = torch.tensor([total], dtype=torch.long, device=device)
71    dist.all_reduce(num_passed_tensor, op=dist.ReduceOp.SUM)
72    dist.all_reduce(total_tensor, op=dist.ReduceOp.SUM)
73    num_passed = num_passed_tensor.item()
74    total = total_tensor.item()
  • 8 lines in 2 placespython76 tokens
    scripts/chat_rl.py:32–39scripts/chat_sft.py:35–42
    32parser = argparse.ArgumentParser(description="Reinforcement learning on GSM8K")
33# Logging
34parser.add_argument("--run", type=str, default="dummy", help="wandb run name ('dummy' disables wandb logging)")
35# Runtime
36parser.add_argument("--device-type", type=str, default="", help="cuda|cpu|mps (empty = autodetect)")
37# Model loading
38parser.add_argument("--model-tag", type=str, default=None, help="model tag to load from")
39parser.add_argument("--model-step", type=int, default=None, help="model step to load from")
  • 10 lines in 2 placesbash73 tokens
    runs/miniseries.sh:72–81runs/scaling_laws.sh:84–93
    72    --model-tag="${TAG}" \
73    --core-metric-every=999999 \
74    --core-metric-max-per-task=-1 \
75    --sample-every=-1 \
76    --save-every=-1 \
77    $DEVICE_BATCH_SIZE_ARG \
78    2>&1 | tee "$RESULTS_DIR/${TAG}_train.log"
79
80END_TIME=$(date +%s)
81TRAIN_TIME=$((END_TIME - START_TIME))
5 more
Largest clones over all files, tests and data included
  • 16 lines in 2 placespythontests268 tokens
    tests/test_attention_fallback.py:160–175tests/test_attention_fallback.py:193–208
    160k_init = torch.randn(B, T_prefill, H, D, device=self.DEVICE, dtype=self.DTYPE)
161v_init = torch.randn(B, T_prefill, H, D, device=self.DEVICE, dtype=self.DTYPE)
162q_single = torch.randn(B, 1, H, D, device=self.DEVICE, dtype=self.DTYPE)
163k_single = torch.randn(B, 1, H, D, device=self.DEVICE, dtype=self.DTYPE)
164v_single = torch.randn(B, 1, H, D, device=self.DEVICE, dtype=self.DTYPE)
165
166def run():
167    k_cache = torch.zeros(B, T_max, H, D, device=self.DEVICE, dtype=self.DTYPE)
168    v_cache = torch.zeros(B, T_max, H, D, device=self.DEVICE, dtype=self.DTYPE)
169    k_cache[:, :T_prefill, :, :] = k_init

    First 10 of 16 lines.

  • 7 lines in 3 placespythontests105 tokens
    tests/test_attention_fallback.py:60–66tests/test_attention_fallback.py:74–80tests/test_attention_fallback.py:89–95
    60B, T, H, D = 2, 64, 4, 32
61q = torch.randn(B, T, H, D, device=self.DEVICE, dtype=self.DTYPE)
62k = torch.randn(B, T, H, D, device=self.DEVICE, dtype=self.DTYPE)
63v = torch.randn(B, T, H, D, device=self.DEVICE, dtype=self.DTYPE)
64
65def run():
66    return flash_attn.flash_attn_func(q, k, v, causal=True, window_size=(T, 0))

    74B, T, H, D = 2, 128, 4, 32
75q = torch.randn(B, T, H, D, device=self.DEVICE, dtype=self.DTYPE)
76k = torch.randn(B, T, H, D, device=self.DEVICE, dtype=self.DTYPE)
77v = torch.randn(B, T, H, D, device=self.DEVICE, dtype=self.DTYPE)
78
79def run():
80    return flash_attn.flash_attn_func(q, k, v, causal=True, window_size=(-1, -1))
  • 6 lines in 3 placestextdocs and examples85 tokens
    dev/LOG.md:505–510dev/LOG.md:517–522dev/LOG.md:529–534
    505Kaplan-style (all projections including lm_head and no embeddings)
506
507Optimal configurations (from quadratic fits):
508FLOPs        Eff Params      Tokens          Ratio      Val BPB
509-----------------------------------------------------------------
5101e+18        110,678,115     1,241,505,403   11.2       0.8972

    517Chinchilla-style (all parameters, period.)
518
519Optimal configurations (from quadratic fits):
520FLOPs        Eff Params      Tokens          Ratio      Val BPB
521-----------------------------------------------------------------
5221e+18        416,320,605     1,232,157,011   3.0        0.8974
  • 6 lines in 2 placespythontests101 tokens
    tests/test_attention_fallback.py:61–66tests/test_attention_fallback.py:121–126
  • 8 lines in 2 placespython86 tokens
    scripts/chat_eval.py:67–74scripts/chat_eval.py:141–148
By format, code only
FormatFilesLinesClonesDuplicated linesDuplication
python296,7157741.1%
bash43841102.6%
By format, all files
FormatFilesLinesClonesDuplicated linesDuplication
python367,810141311.68%
text318474323.37%
bash53921102.55%
markdown41,561000%
toml169000%

Trending appearances

DayRankStarsFilesLinesClonesAll filesCodeHealthCommit
Nov 3, 2026#760,004 (+321)4910,016221.84%1.18%D 42.692d63d4

Want this for your own project? Install jscpd and run jscpd ., or measure it the way the baselines were measured.