The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
T: int64
n_quant_k: int64
n_quant_v: int64
group_size: int64
residual_length: int64
k_bits: int64
v_bits: int64
note: string
format: string
kv_quant: string
limit: int64
lm_eval_task_groups: struct<gsm8k_cot: string, mmlu_cot: string, bbh_cot: string>
child 0, gsm8k_cot: string
child 1, mmlu_cot: string
child 2, bbh_cot: string
prompt_format: string
sample_seed: int64
tasks: list<item: string>
child 0, item: string
dtype: string
model: string
storage_dtype: string
weight_quant: string
to
{'model': Value('string'), 'tasks': List(Value('string')), 'lm_eval_task_groups': {'gsm8k_cot': Value('string'), 'mmlu_cot': Value('string'), 'bbh_cot': Value('string')}, 'limit': Value('int64'), 'sample_seed': Value('int64'), 'prompt_format': Value('string'), 'dtype': Value('string'), 'storage_dtype': Value('string'), 'weight_quant': Value('string'), 'kv_quant': Value('string'), 'format': Value('string')}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2840, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
T: int64
n_quant_k: int64
n_quant_v: int64
group_size: int64
residual_length: int64
k_bits: int64
v_bits: int64
note: string
format: string
kv_quant: string
limit: int64
lm_eval_task_groups: struct<gsm8k_cot: string, mmlu_cot: string, bbh_cot: string>
child 0, gsm8k_cot: string
child 1, mmlu_cot: string
child 2, bbh_cot: string
prompt_format: string
sample_seed: int64
tasks: list<item: string>
child 0, item: string
dtype: string
model: string
storage_dtype: string
weight_quant: string
to
{'model': Value('string'), 'tasks': List(Value('string')), 'lm_eval_task_groups': {'gsm8k_cot': Value('string'), 'mmlu_cot': Value('string'), 'bbh_cot': Value('string')}, 'limit': Value('int64'), 'sample_seed': Value('int64'), 'prompt_format': Value('string'), 'dtype': Value('string'), 'storage_dtype': Value('string'), 'weight_quant': Value('string'), 'kv_quant': Value('string'), 'format': Value('string')}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
FP16/BF16 and SmoothQuant W8A8 + KIVI-INT8 KV (Llama-3.1-8B-Instruct, CoT)
This dataset combines FP16, BF16, and SmoothQuant W8A8 + KIVI-INT8 reference artifacts for meta-llama/Llama-3.1-8B-Instruct.
It follows the artifact layouts of:
There are three model/KV configurations and three CoT task groups, giving nine task/configuration KV combinations. Weight artifacts are stored once per model configuration.
| Configuration | Weight | Activation | KV cache |
|---|---|---|---|
| W-FP16/A-FP16/KV-FP16 | FP16 | FP16 | FP16 |
| W-BF16/A-BF16/KV-BF16 | BF16 | BF16 | BF16 |
| SmoothQuant W8A8/KIVI-KV8 | per-channel symmetric INT8 | per-token symmetric INT8 | asymmetric KIVI INT8 |
w_of_wfp16a16kvfp16_llama_31_8b/ - FP16 weights
The FP16 weights of Llama-3.1-8B-Instruct.
Stored per layer: layer_0.safetensors ... layer_31.safetensors plus embeddings.safetensors and config.json.
The seven linears per layer are stored as FP16 tensors; layer norms, embeddings, final norm, and LM head are also FP16:
| key | dtype | shape |
|---|---|---|
self_attn.q_proj.weight, self_attn.o_proj.weight |
fp16 | (4096, 4096) |
self_attn.k_proj.weight, self_attn.v_proj.weight |
fp16 | (1024, 4096), GQA (8 KV heads x 128) |
mlp.gate_proj.weight, mlp.up_proj.weight |
fp16 | (14336, 4096) |
mlp.down_proj.weight |
fp16 | (4096, 14336) |
*.scale (identity scale) |
fp32 | (1,) |
input_layernorm.weight, post_attention_layernorm.weight |
fp16 | (4096,) |
model.embed_tokens.weight, lm_head.weight |
fp16 | (128256, 4096) |
model.norm.weight |
fp16 | (4096,) |
w_of_wbf16abf16kvbf16_llama_31_8b/ - BF16 weights
The BF16 weights of Llama-3.1-8B-Instruct.
Stored per layer: layer_0.safetensors ... layer_31.safetensors plus embeddings.safetensors and config.json.
The seven linears per layer are stored as BF16 tensors; layer norms, embeddings, final norm, and LM head are also BF16:
| key | dtype | shape |
|---|---|---|
self_attn.q_proj.weight, self_attn.o_proj.weight |
bf16 | (4096, 4096) |
self_attn.k_proj.weight, self_attn.v_proj.weight |
bf16 | (1024, 4096), GQA (8 KV heads x 128) |
mlp.gate_proj.weight, mlp.up_proj.weight |
bf16 | (14336, 4096) |
mlp.down_proj.weight |
bf16 | (4096, 14336) |
*.scale (identity scale) |
fp32 | (1,) |
input_layernorm.weight, post_attention_layernorm.weight |
bf16 | (4096,) |
model.embed_tokens.weight, lm_head.weight |
bf16 | (128256, 4096) |
model.norm.weight |
bf16 | (4096,) |
w8_of_w8a8_smoothquant_llama_31_8b/ - SmoothQuant W8A8 weights
The INT8 weights of the SmoothQuant W8A8 model. SmoothQuant uses alpha=0.85, per-output-channel symmetric weight quantization, and dynamic per-token symmetric activation quantization. Activation quantization is applied during inference and is not stored as a separate artifact.
The activation scale file used for smoothing is llama-3.1-8b-instruct.pt. The W8A8 model is constructed by applying smooth_lm followed by SmoothQuant quantize_model(weight_quant="per_channel", act_quant="per_token", quantize_bmm_input=False).
Stored per layer: layer_0.safetensors ... layer_31.safetensors plus embeddings.safetensors and config.json.
The seven linears per layer are quantized as INT8 weights with FP16 per-output-channel scales; everything else stays FP16. Dequantization is weight_fp = weight_int8 * scale.
| key | dtype | shape |
|---|---|---|
self_attn.q_proj.weight, self_attn.o_proj.weight |
int8 | (4096, 4096) |
self_attn.k_proj.weight, self_attn.v_proj.weight |
int8 | (1024, 4096), GQA (8 KV heads x 128) |
mlp.gate_proj.weight, mlp.up_proj.weight |
int8 | (14336, 4096) |
mlp.down_proj.weight |
int8 | (4096, 14336) |
*.scale (per quantized linear) |
fp16 | (out_features, 1) |
input_layernorm.weight, post_attention_layernorm.weight |
fp16 | (4096,) |
model.embed_tokens.weight, lm_head.weight |
fp16 | (128256, 4096) |
model.norm.weight |
fp16 | (4096,) |
kv_fp16_of_wfp16a16kvfp16_llama_31_8b/ - FP16 KV cache
The post-prefill KV cache of the FP16 model, stored directly as FP16 tensors.
Layout: <task>/sample_<n>.safetensors, with all 32 layers in one file. File metadata contains T (sequence length), dtype=fp16, storage_dtype=float16, and n_layers=32.
| key | dtype | shape (T = sequence length) |
|---|---|---|
layer_<i>.k_code |
fp16 | (1, 8, T, 128) |
layer_<i>.k_scale |
fp32 | (1,), identity scale |
layer_<i>.v_code |
fp16 | (1, 8, T, 128) |
layer_<i>.v_scale |
fp32 | (1,), identity scale |
kv_bf16_of_wbf16abf16kvbf16_llama_31_8b/ - BF16 KV cache
The post-prefill KV cache of the BF16 model, stored directly as BF16 tensors.
Layout: <task>/sample_<n>.safetensors, with all 32 layers in one file. File metadata contains T (sequence length), dtype=bf16, storage_dtype=bfloat16, and n_layers=32.
| key | dtype | shape (T = sequence length) |
|---|---|---|
layer_<i>.k_code |
bf16 | (1, 8, T, 128) |
layer_<i>.k_scale |
fp32 | (1,), identity scale |
layer_<i>.v_code |
bf16 | (1, 8, T, 128) |
layer_<i>.v_scale |
fp32 | (1,), identity scale |
kv_kivi8_of_w8a8_smoothquant_llama_31_8b/ - KIVI-INT8 KV cache
The post-prefill KV cache of the same SmoothQuant W8A8 model, quantized with KIVI INT8: asymmetric Key per-channel and Value per-token quantization with group_size=32, residual_length=128, k_bits=8, and v_bits=8.
Layout: <task>/sample_<n>/layer_<i>.safetensors. Each layer file contains {k_code, k_scale, k_min, v_code, v_scale, v_min}. Codes are packed INT32 tensors; scale and minimum tensors are FP16.
The snapshot is taken immediately after prefill and before generation. Only the packed INT8 region is stored; the latest 128-token FP16 residual is intentionally excluded, matching the reference layout.
| key | dtype | shape (T = sequence length) |
|---|---|---|
k_code |
int32 | (1, 8, n_quant_k/4, 128) |
k_scale, k_min |
fp16 | (1, 8, n_quant_k/32, 1, 128) |
v_code |
int32 | (1, 8, n_quant_v, 128/4) |
v_scale, v_min |
fp16 | (1, 8, n_quant_v, 128/32, 1) |
n_quant_k = floor((T - 128) / 32) * 32 and n_quant_v = T - 128.
Tasks and prompts
Each KV-cache configuration contains 20 deterministically sampled prompts from each task group:
| Dataset directory | lm-eval task/group | Samples |
|---|---|---|
gsm8k_cot |
gsm8k_cot |
20 |
mmlu_cot |
mmlu_cot_llama |
20 |
bbh_cot |
bbh_cot_fewshot |
20 |
Prompts were built with lm-eval 0.4.11 using the Llama-3.1-Instruct chat template, apply_chat_template=True, and fewshot_as_multiturn=True. The task-configured few-shot settings are retained. manifests/prompts.jsonl records the exact prompt text, selected document ID, target, generation arguments, and token count for every sample.
All KV files are post-prefill snapshots. No answer tokens are generated or included in the cache.
Validation
manifests/complete.json records the completed artifact counts. The dataset contains:
- 34 weight files per configuration: 32 layer files,
embeddings.safetensors, andconfig.json - 20 KV samples per task and configuration
- 180 KV samples total across the nine task/configuration combinations
All weight and KV artifacts were validated for required files, tensor keys, dtype, shape, sequence-length metadata, prompt-manifest consistency, and KIVI packing layout.
- Downloads last month
- 608