Dataset Viewer
Auto-converted to Parquet Duplicate
text
stringlengths
0
1.33k
13397 MiB
13339 MiB
13295 MiB
11365 MiB
0 MiB
0 MiB
0 MiB
0 MiB
== prompt
{"prompt":"<|hy_start:opensource|>system<|hy_middle:opensource|># Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>\n{\"name\": \"get_weather\", \"description\": \"Current weather for a city\", \"parameters\"...
== port 8089 no_think
{"choices":[{"finish_reason":"tool_calls","index":0,"message":{"role":"assistant","content":"","tool_calls":[{"type":"function","function":{"name":"get_weather","arguments":"{\"city\":\"Paris\"}"},"id":"X5m8PjvwWuHEtDK62i277oXHzD8DTqS1"}]}}],"created":1795008211,"model":"/work/q8/Hy4-flash-Q8_0-00001-of-00002.gguf","sy...
== port 8089 high
{"choices":[{"finish_reason":"tool_calls","index":0,"message":{"role":"assistant","content":"","tool_calls":[{"type":"function","function":{"name":"get_weather","arguments":"{\"city\":\"Paris\"}"},"id":"eE4vwLNpY1YlPz4MktnSM61CvMf3Ymtm"}]}}],"created":1795008213,"model":"/work/q8/Hy4-flash-Q8_0-00001-of-00002.gguf","sy...
== port 8088 no_think
{"choices":[{"finish_reason":"stop","index":0,"message":{"role":"assistant","content":"I don't have access to real-time weather data, so I can't check the current weather in Paris directly. You can use reliable sources like:\n\n- **Weather apps/websites**: AccuWeather, Weather.com, or local services like Météo-France (...
== port 8088 high
{"choices":[{"finish_reason":"tool_calls","index":0,"message":{"role":"assistant","content":"","reasoning_content":"The user is asking about the weather in Paris. I need to use the weather tool to get the current weather for Paris. The tool takes a city parameter. I'll call get_weather with city=\"Paris\".","tool_calls...
TOOL_DONE
/work/results/imatrix.gguf: 1114 chunks
814 entries, 123 with experts
dead experts (count 0): 0 tensor-expert pairs, 0 distinct (block, expert); expected by the router 0, unexpected 0
weak experts per block: 23:1, 37:1, 40:1
weak experts (< 1% of the median count): 9
lowest min/median:
blk.40.ffn_down_exps.weight min 69 median 11066 max 243786
blk.40.ffn_gate_exps.weight min 69 median 11066 max 243786
blk.40.ffn_up_exps.weight min 69 median 11066 max 243786
blk.23.ffn_down_exps.weight min 74 median 10571 max 194219
blk.23.ffn_gate_exps.weight min 74 median 10571 max 194219
blk.23.ffn_up_exps.weight min 74 median 10571 max 194219
blk.37.ffn_down_exps.weight min 105 median 11394 max 249909
blk.37.ffn_gate_exps.weight min 105 median 11394 max 249909
quantisable weights with no imatrix entry: 29.28% of parameters
per_layer_token_embd.weight 1 tensors 28.94% GET_ROWS table, no matmul to observe
output.weight 1 tensors 0.17% llama-imatrix skips it without --process-output
token_embd.weight 1 tensors 0.17% GET_ROWS table, no matmul to observe
blk.N.engram_kv_b.weight 4 tensors 0.00% a blk.* matmul with no entry: never executed (MTP?) or the corpus missed it
output_hc_fn.weight 1 tensors 0.00% outside blk.*, llama-imatrix never collects it
blk.N.engram_gate.weight 4 tensors 0.00% a blk.* matmul with no entry: never executed (MTP?) or the corpus missed it
blk.N.engram_hc_scale.weight 4 tensors 0.00% a blk.* matmul with no entry: never executed (MTP?) or the corpus missed it
blk.N.indexer_compressor_ape.weight 12 tensors 0.00% a blk.* matmul with no entry: never executed (MTP?) or the corpus missed it
0.04.206.636 W compute_tensor_statistics: no preceding-layer tensor for 'blk.1.ffn_gate_inp.weight'
0.04.218.778 W compute_tensor_statistics: no preceding-layer tensor for 'blk.1.ffn_up_exps.weight'
0.04.255.661 W compute_tensor_statistics: no preceding-layer tensor for 'blk.1.ffn_routed_up.weight'
0.04.270.794 W compute_tensor_statistics: no preceding-layer tensor for 'blk.1.ffn_gate_exps.weight'
0.04.274.000 W compute_tensor_statistics: no preceding-layer tensor for 'blk.1.ffn_down_exps.weight'
0.04.281.859 W compute_tensor_statistics: no preceding-layer tensor for 'blk.1.ffn_down_shexp.weight'
0.04.308.919 W compute_tensor_statistics: no preceding-layer tensor for 'blk.1.ffn_up_shexp.weight'
0.04.324.368 W compute_tensor_statistics: no preceding-layer tensor for 'blk.1.engram_kv.weight'
0.04.431.941 W compute_tensor_statistics: no preceding-layer tensor for 'blk.1.ffn_routed_down.weight'
0.04.452.430 W compute_tensor_statistics: no preceding-layer tensor for 'blk.1.ffn_gate_shexp.weight'
Computing tensor statistics for /work/results/imatrix.gguf (814 tensors)
Layer | Tensor | Mean StdDev Skew Kurt H Norm | ∑ E[A²] Gain | L2 Dist PCC Cov
---------------------------------------------------------------------------------------------------------------------------------------------------------------------
0 | blk.0.attn_gate.weight | 0.0014 0.0743 -0.0150 0.0263 98.95% | 59.7349 nan | nan nan nan
0 | blk.0.attn_k_b.weight | -0.0037 0.9850 0.0232 2.6646 94.05% | 9804.2355 nan | nan nan nan
0 | blk.0.attn_kv_a_mqa.weight | 0.0014 0.0743 -0.0150 0.0263 98.95% | 59.7349 nan | nan nan nan
0 | blk.0.attn_output.weight | -0.0003 0.0451 0.0206 -0.7867 95.04% | 17.6732 nan | nan nan nan
0 | blk.0.attn_q_a.weight | 0.0014 0.0743 -0.0150 0.0263 98.95% | 59.7349 nan | nan nan nan
0 | blk.0.attn_q_b.weight | 0.0021 0.2845 1.1939 30.9028 84.28% | 168.9323 nan | nan nan nan
0 | blk.0.attn_v_b.weight | -0.0001 0.1066 -0.1111 12.9396 78.86% | 193.4295 nan | nan nan nan
0 | blk.0.ffn_down.weight | 0.0000 0.0003 -0.0663 4.9242 95.10% | 0.0082 nan | nan nan nan
0 | blk.0.ffn_gate.weight | -0.0008 0.0315 -32.6422 1207.8180 39.35% | 3.7587 nan | nan nan nan
0 | blk.0.ffn_up.weight | -0.0008 0.0315 -32.6422 1207.8180 39.35% | 3.7587 nan | nan nan nan
0 | blk.0.hc_attn_fn.weight | 0.0120 0.5285 -0.0400 -0.4391 99.18% | 12196.4277 nan | nan nan nan
0 | blk.0.hc_ffn_fn.weight | -0.0094 0.6133 -2.7317 45.9505 95.05% | 12152.2965 nan | nan nan nan
0 | blk.0.indexer_compressor_gate.weight | -0.0434 0.6365 -0.0552 -0.1210 95.76% | 99.6459 nan | nan nan nan
0 | blk.0.indexer_compressor_value.weight | -0.0434 0.6365 -0.0552 -0.1210 95.76% | 99.6459 nan | nan nan nan
0 | blk.0.indexer.attn_k.weight | 0.0014 0.0743 -0.0150 0.0263 98.95% | 59.7349 nan | nan nan nan
0 | blk.0.indexer.attn_q_b.weight | 0.0021 0.2845 1.1939 30.9028 84.28% | 168.9323 nan | nan nan nan
0 | blk.0.indexer.proj.weight | 0.0014 0.0743 -0.0150 0.0263 98.95% | 59.7349 nan | nan nan nan
1 | blk.1.attn_gate.weight | 0.0013 0.1109 -3.2728 58.6003 92.28% | 87.3179 1.21 | 6.9494 0.1268 0.0010
1 | blk.1.attn_k_b.weight | 0.0020 0.6556 -0.0938 1.4539 94.68% | 4381.4724 0.67 | 92.5342 0.0049 0.0032
1 | blk.1.attn_kv_a_mqa.weight | 0.0013 0.1109 -3.2728 58.6003 92.28% | 87.3179 1.21 | 6.9494 0.1268 0.0010
1 | blk.1.attn_output.weight | 0.0001 0.0060 0.0291 9.3798 92.77% | 1.3470 0.28 | 4.1150 0.0117 0.0000
1 | blk.1.attn_q_a.weight | 0.0013 0.1109 -3.2728 58.6003 92.28% | 87.3179 1.21 | 6.9494 0.1268 0.0010
1 | blk.1.attn_q_b.weight | 0.0033 0.1993 0.2671 8.6021 93.55% | 119.8413 0.84 | 11.0965 0.0037 0.0002
1 | blk.1.attn_v_b.weight | -0.0007 0.0298 -0.0462 0.5541 95.92% | 75.0937 0.62 | 14.1417 0.0085 0.0000
1 | blk.1.engram_kv.weight | -0.0029 0.1212 0.0393 -0.0644 99.80% | 492.6544 nan | nan nan nan
1 | blk.1.ffn_down_exps.weight | 0.0000 0.0006 10.3639 1865.2014 87.36% | 2.3984 nan | nan nan nan
1 | blk.1.ffn_down_shexp.weight | -0.0000 0.0037 -0.3403 12.8903 91.71% | 0.0766 nan | nan nan nan
1 | blk.1.ffn_gate_exps.weight | 0.0006 0.0186 0.0010 0.1293 99.81% | 2265.6961 nan | nan nan nan
1 | blk.1.ffn_gate_inp.weight | 0.0011 0.0266 0.2603 15.7296 86.38% | 7.2478 nan | nan nan nan
1 | blk.1.ffn_gate_shexp.weight | 0.0011 0.0266 0.2603 15.7296 86.38% | 7.2478 nan | nan nan nan
1 | blk.1.ffn_routed_down.weight | 0.0011 0.0266 0.2603 15.7296 86.38% | 7.2478 nan | nan nan nan
1 | blk.1.ffn_routed_up.weight | -0.0000 0.0009 -0.0841 -0.1298 99.44% | 0.0132 nan | nan nan nan
1 | blk.1.ffn_up_exps.weight | 0.0006 0.0186 0.0010 0.1293 99.81% | 2265.6961 nan | nan nan nan
1 | blk.1.ffn_up_shexp.weight | 0.0011 0.0266 0.2603 15.7296 86.38% | 7.2478 nan | nan nan nan
1 | blk.1.hc_attn_fn.weight | 0.0022 0.6161 -1.3003 19.1612 95.48% | 12194.0913 1.00 | 85.7207 0.0938 0.0305
1 | blk.1.hc_ffn_fn.weight | 0.0047 0.6016 -1.3391 21.0434 95.51% | 12197.9626 1.00 | 42.1830 0.8042 0.2967
1 | blk.1.indexer_compressor_gate.weight | -0.0367 0.7870 -0.3461 1.1651 89.83% | 109.1556 1.05 | 11.7163 -0.0478 -0.0239
1 | blk.1.indexer_compressor_value.weight | -0.0367 0.7870 -0.3461 1.1651 89.83% | 109.1556 1.05 | 11.7163 -0.0478 -0.0239
1 | blk.1.indexer.attn_k.weight | 0.0013 0.1109 -3.2728 58.6003 92.28% | 87.3179 1.21 | 6.9494 0.1268 0.0010
1 | blk.1.indexer.attn_q_b.weight | 0.0033 0.1993 0.2671 8.6021 93.55% | 119.8413 0.84 | 11.0965 0.0037 0.0002
1 | blk.1.indexer.proj.weight | 0.0013 0.1109 -3.2728 58.6003 92.28% | 87.3179 1.21 | 6.9494 0.1268 0.0010
2 | blk.2.attn_gate.weight | 0.0022 0.1281 -15.1570 564.4596 83.38% | 113.7723 1.14 | 5.0494 0.7183 0.0102
2 | blk.2.attn_k_b.weight | 0.0170 0.6273 -0.0126 1.3304 95.45% | 8246.3797 nan | nan nan nan
End of preview. Expand in Data Studio

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Hy4-flash GGUF metrics

Measurements behind AtomicChat/Hy4-flash-GGUF.

  • kld/: llama-perplexity --kl-divergence logs on the 30-language neutral corpus (104 chunks of 4096 tokens), against our BF16 GGUF of tencent/Hy4-flash (rev d88c9dfd). llama.cpp PR #32517 at 0a054d0b, 8x A100 80 GB. base-neutral.log is the reference run and base-selfcheck.log the reference against itself (KLD 0).
    • kld-AD-IQ2_XXS-IQ2_S.log is the published file (named AD-IQ2_S-IQ2_XXS there: down IQ2_S, gate/up IQ2_XXS).
    • kld-vendor-*.log are Tencent's AngelSlim files on the same stand.
    • -T8 / -T8_0 are the same weights with the Engram table at F8_E4M3 / Q8_0.
  • imatrix/: the importance matrix (imatrix.gguf) with its statistics, expert coverage and log. It was computed on the BF16 over 1,114 chunks of 512 tokens: an even eighth of the hy4-flash-moe calibration build. Rebuild the input with subset.py calib_train.txt calib_sub.txt --pieces 8901 --every 8 (pipeline/experiments/minimax31/subset.py).
  • ladder/: the per-variant --tensor-type rules and the profile they came from.
  • logs/: conversion and quantization logs and the tensor-type verification.
  • checks/:
    • validate.json: our BF16 against Tencent's GGUF (Engram token map, metadata, byte comparison);
    • smoke.jsonl / tooltest.txt: hand checks (chat, reasoning, Chinese, code, image, tool calls);
    • smoke-mem.txt: GPU memory with the published file.
Downloads last month
106

Collection including AtomicChat/Hy4-flash-GGUF-metrics