KeelCore-750M-Friendly / adapter_config.json
saravanan-arjuna-ravi's picture
Update to the phase-2 adapter: dates, numbers and exact text are kept more reliably
d5a7563 verified
Raw History Blame Contribute Delete
3.96 kB
{
"format_version": 1,
"adapter": {
"rank": 16,
"alpha": 8.0,
"use_rslora": true,
"dropout": 0.05,
"targets": {
"encoder_self_attention": [
"q_proj",
"k_proj",
"v_proj",
"o_proj"
],
"encoder_ffn": [
"gate_proj",
"up_proj",
"down_proj"
],
"decoder_self_attention": [
"q_proj",
"k_proj",
"v_proj",
"o_proj"
],
"decoder_cross_attention": [
"q_proj",
"k_proj",
"v_proj",
"o_proj"
],
"decoder_ffn": [
"gate_proj",
"up_proj",
"down_proj"
]
},
"lr_multipliers": {},
"trainable_tokens": [
"<style:friendly>",
"<text_to_rewrite>",
"</text_to_rewrite>"
]
},
"scaling": 2.0,
"style_token": "<style:friendly>",
"trainable_tokens": {
"<style:friendly>": 266,
"<text_to_rewrite>": 274,
"</text_to_rewrite>": 275
},
"parameters": 18287616,
"base": {
"checkpoint": "KeelCore/KeelCore-750M",
"step": 90000,
"fingerprint": "3dcb3cf25d47379e86e12d889e9139d59cf7508846b0bd036b9fb97200968cd7"
},
"provenance": {
"phase_1": {
"step": 146342,
"final_metrics": {
"eval/loss": 0.2977872219764001,
"eval/perplexity": 1.3468751719507266,
"eval/token_accuracy": 0.9084640896078094,
"eval/eos_accuracy": 0.9791729398775469,
"eval/ece": 0.004893604238046363,
"eval/confidence": 0.9129432059253049,
"eval/examples": 19926.0,
"eval/loss_by_length/short": 0.34995758793783915,
"eval/loss_by_length/medium": 0.3643175207048626,
"eval/loss_by_length/long": 0.2319568483194215,
"generation/sari": 65.33603310642951,
"generation/chrf_reference": 77.19320224188608,
"generation/chrf_input": 78.11275879064725,
"generation/length_ratio": 1.0316882908212353,
"generation/shorter_rate": 0.364,
"generation/span_preservation": 0.951260678959641,
"generation/novel_word_rate": 0.16575745373099998,
"generation/loop_rate": 0.002,
"generation/terminated_rate": 0.982,
"generation/empty_rate": 0.0,
"generation/copy_rate": 0.004,
"generation/distinct_2": 0.7101879672978827,
"generation/examples": 500.0,
"generation/output_tokens": 84.684,
"quality/acceptable_rate": 0.93,
"perf/eval_seconds": 158.31764645129442
}
},
"phase_2": {
"step": 1444,
"final_metrics": {
"eval/loss": 0.31521833835457524,
"eval/perplexity": 1.370558523783922,
"eval/token_accuracy": 0.9076035833559555,
"eval/eos_accuracy": 0.9835221421215242,
"eval/ece": 0.006463088998026999,
"eval/confidence": 0.9128969446645174,
"eval/examples": 3884.0,
"eval/loss_by_length/short": 0.3533447977725127,
"eval/loss_by_length/medium": 0.3540014474260306,
"eval/loss_by_length/long": 0.2640006830703098,
"generation/sari": 67.20927204646436,
"generation/chrf_reference": 78.57761231766997,
"generation/chrf_input": 75.47220212143709,
"generation/length_ratio": 1.0527019284083794,
"generation/shorter_rate": 0.276,
"generation/span_preservation": 0.9128116732687585,
"generation/novel_word_rate": 0.1772604663624468,
"generation/loop_rate": 0.004,
"generation/terminated_rate": 0.984,
"generation/empty_rate": 0.0,
"generation/copy_rate": 0.004,
"generation/distinct_2": 0.7617565439135978,
"generation/examples": 500.0,
"generation/output_tokens": 81.546,
"quality/acceptable_rate": 0.834,
"perf/eval_seconds": 76.68794807884842
},
"note": "continued from the phase-1 adapter on ~20k date/link-rich rewrites mixed 1:1 with clean phase-1 rows; 1 epoch, lr 5e-5"
}
}
}