Activation capture run configuration
Browse files- run_config.json +83 -0
run_config.json
ADDED
|
@@ -0,0 +1,83 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model": "Qwen/Qwen3.5-2B",
|
| 3 |
+
"output_repo": "saicr/Full-Model-Background-Processing",
|
| 4 |
+
"total_sequences": 1000000,
|
| 5 |
+
"max_input_tokens": 128,
|
| 6 |
+
"min_input_tokens_target": 48,
|
| 7 |
+
"max_new_tokens": 16,
|
| 8 |
+
"batch_size": 16,
|
| 9 |
+
"shard_sequences": 128,
|
| 10 |
+
"datasets": [
|
| 11 |
+
{
|
| 12 |
+
"short_name": "ultrachat",
|
| 13 |
+
"repo": "HuggingFaceH4/ultrachat_200k",
|
| 14 |
+
"config": null,
|
| 15 |
+
"split": "train_sft",
|
| 16 |
+
"field_mode": "prompt"
|
| 17 |
+
},
|
| 18 |
+
{
|
| 19 |
+
"short_name": "fineweb_edu",
|
| 20 |
+
"repo": "HuggingFaceFW/fineweb-edu",
|
| 21 |
+
"config": "sample-10BT",
|
| 22 |
+
"split": "train",
|
| 23 |
+
"field_mode": "text"
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"short_name": "openwebmath",
|
| 27 |
+
"repo": "open-web-math/open-web-math",
|
| 28 |
+
"config": null,
|
| 29 |
+
"split": "train",
|
| 30 |
+
"field_mode": "text"
|
| 31 |
+
},
|
| 32 |
+
{
|
| 33 |
+
"short_name": "cosmopedia_v2",
|
| 34 |
+
"repo": "HuggingFaceTB/smollm-corpus",
|
| 35 |
+
"config": "cosmopedia-v2",
|
| 36 |
+
"split": "train",
|
| 37 |
+
"field_mode": "cosmopedia"
|
| 38 |
+
}
|
| 39 |
+
],
|
| 40 |
+
"formats": [
|
| 41 |
+
"raw",
|
| 42 |
+
"qa_short",
|
| 43 |
+
"user_assistant",
|
| 44 |
+
"hash_sections",
|
| 45 |
+
"xml_document",
|
| 46 |
+
"key_point",
|
| 47 |
+
"continue",
|
| 48 |
+
"next_sentence",
|
| 49 |
+
"context_answer",
|
| 50 |
+
"brackets",
|
| 51 |
+
"reader_reply",
|
| 52 |
+
"plain_instruction",
|
| 53 |
+
"title",
|
| 54 |
+
"summary",
|
| 55 |
+
"minimal",
|
| 56 |
+
"note",
|
| 57 |
+
"completion",
|
| 58 |
+
"quoted",
|
| 59 |
+
"dash",
|
| 60 |
+
"task_block"
|
| 61 |
+
],
|
| 62 |
+
"sampling": {
|
| 63 |
+
"temperature": 0.85,
|
| 64 |
+
"top_k": 50,
|
| 65 |
+
"top_p": 0.95
|
| 66 |
+
},
|
| 67 |
+
"capture": {
|
| 68 |
+
"meaning": "Transformers hidden_states tuple: embedding/residual stream plus all model layers",
|
| 69 |
+
"hidden_size": 2048,
|
| 70 |
+
"num_hidden_layers": 24,
|
| 71 |
+
"num_hidden_state_tensors": 25,
|
| 72 |
+
"prefill": true,
|
| 73 |
+
"generated_tokens": true
|
| 74 |
+
},
|
| 75 |
+
"quantization": {
|
| 76 |
+
"dtype": "int8",
|
| 77 |
+
"scheme": "symmetric per-token per-hidden-state",
|
| 78 |
+
"scale_dtype": "float16",
|
| 79 |
+
"formula": "float ~= int8 * scale"
|
| 80 |
+
},
|
| 81 |
+
"tokenizer_class": "Qwen2Tokenizer",
|
| 82 |
+
"created_unix": 1792240935.4077983
|
| 83 |
+
}
|