Dataset Viewer
Auto-converted to Parquet Duplicate
adapter_role
string
base_model
string
parameter_cap
int64
total_parameters_with_adapter
int64
trainable_parameters
int64
lora_targets
list
system_prompt
string
mapping
dict
split_protocol
string
training_targets
string
max_steps_planned
int64
max_steps_completed
int64
elapsed_hours
float64
eval_metrics
dict
evaluation_protocol
string
run_sizing
dict
config
dict
source_translation
unsloth/gemma-2-2b-it
3,000,000,000
2,655,875,328
41,533,440
[ "q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj" ]
Translate the supplied English doctor answer into natural Bengali. Preserve every medical fact, medicine, test, dosage, uncertainty, warning, and sentence order. Match the Bengali style used by Nascenia training responses. Do not diagnose independently and do not add information absent from the English source answer. R...
{ "rows": 112165, "external_rows": 112156, "competition_train_rows": 108954, "competition_test_rows": 1000, "mapped_train_rows": 108945, "mapped_test_rows": 1000, "missing_test_ids": [], "source_gap_ids": [ 219, 16222, 19056, 27951, 66346, 81014, 84575, 90042, 105487 ...
English-source-answer grouped holdout
Competition bn_response only; zero synthetic targets
2,066
2,066
9.725501
{ "eval_loss": 0.18040531873703003, "eval_runtime": 2011.6946, "eval_samples_per_second": 2.716, "eval_steps_per_second": 1.358, "epoch": 0.3263629721777936 }
No periodic evaluation; one final loss evaluation
{ "seconds_per_step": 16.46638858318329, "full_epoch_steps": 6331, "budget_steps": 2066, "selected_max_steps": 2066, "examples_seen": 33056, "epoch_fraction": 0.32633075343547624 }
{ "base_model": "unsloth/gemma-2-2b-it", "adapter_name": "gemma2_source_translation_lora", "adapter_role": "source_translation", "parameter_cap": 3000000000, "seed": 42, "max_seq_length": 1024, "max_prompt_tokens": 480, "load_in_4bit": false, "per_device_batch_size": 2, "gradient_accumulation_steps"...
README.md exists but content is empty.
Downloads last month
14