entrenar 0.7.9

Training & Optimization library with autograd, LoRA, quantization, and model merging
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231

//! Falsification tests for HF leaderboard and publishing integration
//!
//! Tests the contracts between leaderboard parsing, model card generation,
//! and submission formatting against HuggingFace Hub specifications.

use crate::eval::evaluator::{EvalResult, Leaderboard, Metric};
use crate::eval::RougeVariant;

use super::parser::{column_to_metric, compare_with_leaderboard, to_leaderboard};
use super::types::{HfLeaderboard, LeaderboardEntry, LeaderboardKind};

// ═══════════════════════════════════════════════════════════════════════
// CLAIM: Leaderboard ranking respects higher_is_better semantics
// ═══════════════════════════════════════════════════════════════════════

#[test]
fn falsify_wer_leaderboard_ranks_lower_first() {
    // WER is lower-is-better: model with WER=0.05 should beat WER=0.10
    let mut hf = HfLeaderboard::new(LeaderboardKind::OpenASR);

    let mut bad = LeaderboardEntry::new("bad-model");
    bad.scores.insert("wer".into(), 0.20);
    hf.entries.push(bad);

    let mut good = LeaderboardEntry::new("good-model");
    good.scores.insert("wer".into(), 0.05);
    hf.entries.push(good);

    let lb = to_leaderboard(&hf);
    assert_eq!(
        lb.best().expect("operation should succeed").model_name,
        "good-model",
        "WER leaderboard should rank lower WER first"
    );
}

#[test]
fn falsify_bleu_leaderboard_ranks_higher_first() {
    let mut lb = Leaderboard::new(Metric::BLEU);

    let mut low = EvalResult::new("low-bleu");
    low.add_score(Metric::BLEU, 0.2);
    lb.add(low);

    let mut high = EvalResult::new("high-bleu");
    high.add_score(Metric::BLEU, 0.8);
    lb.add(high);

    assert_eq!(
        lb.best().expect("operation should succeed").model_name,
        "high-bleu",
        "BLEU leaderboard should rank higher BLEU first"
    );
}

#[test]
fn falsify_perplexity_leaderboard_ranks_lower_first() {
    let mut lb = Leaderboard::new(Metric::Perplexity);

    let mut high_ppl = EvalResult::new("high-ppl");
    high_ppl.add_score(Metric::Perplexity, 100.0);
    lb.add(high_ppl);

    let mut low_ppl = EvalResult::new("low-ppl");
    low_ppl.add_score(Metric::Perplexity, 5.0);
    lb.add(low_ppl);

    assert_eq!(
        lb.best().expect("operation should succeed").model_name,
        "low-ppl",
        "Perplexity leaderboard should rank lower PPL first"
    );
}

// ═══════════════════════════════════════════════════════════════════════
// CLAIM: compare_with_leaderboard correctly inserts user model
// ═══════════════════════════════════════════════════════════════════════

#[test]
fn falsify_compare_preserves_all_entries() {
    let mut hf = HfLeaderboard::new(LeaderboardKind::OpenASR);
    for i in 0..5 {
        let mut entry = LeaderboardEntry::new(format!("model-{i}"));
        entry.scores.insert("wer".into(), 0.1 * f64::from(i + 1));
        hf.entries.push(entry);
    }

    let mut my = EvalResult::new("my-model");
    my.add_score(Metric::WER, 0.07);

    let lb = compare_with_leaderboard(&my, &hf);
    assert_eq!(lb.results.len(), 6, "Should have 5 HF entries + 1 user entry");
    assert_eq!(
        lb.best().expect("operation should succeed").model_name,
        "my-model",
        "User model with WER=0.07 should be best"
    );
}

#[test]
fn falsify_compare_user_model_worse_than_all() {
    let mut hf = HfLeaderboard::new(LeaderboardKind::OpenASR);
    let mut entry = LeaderboardEntry::new("sota-model");
    entry.scores.insert("wer".into(), 0.01);
    hf.entries.push(entry);

    let mut my = EvalResult::new("my-bad-model");
    my.add_score(Metric::WER, 0.50);

    let lb = compare_with_leaderboard(&my, &hf);
    assert_eq!(
        lb.best().expect("operation should succeed").model_name,
        "sota-model",
        "SOTA model should remain best"
    );
    // User model should be last
    assert_eq!(
        lb.results.last().expect("collection should not be empty").model_name,
        "my-bad-model"
    );
}

// ═══════════════════════════════════════════════════════════════════════
// CLAIM: column_to_metric handles case insensitivity
// ═══════════════════════════════════════════════════════════════════════

#[test]
fn falsify_column_mapping_case_insensitive() {
    // The spec says column names should be lowercased before matching
    assert_eq!(column_to_metric(&LeaderboardKind::OpenASR, "WER"), Some(Metric::WER));
    assert_eq!(column_to_metric(&LeaderboardKind::OpenASR, "Wer"), Some(Metric::WER));
    assert_eq!(column_to_metric(&LeaderboardKind::OpenLLMv2, "MMLU"), Some(Metric::MMLUAccuracy));
    assert_eq!(
        column_to_metric(&LeaderboardKind::BigCodeBench, "Pass@1"),
        Some(Metric::PassAtK(1))
    );
}

#[test]
fn falsify_column_mapping_unknown_returns_none() {
    assert_eq!(column_to_metric(&LeaderboardKind::OpenASR, "completely_unknown_metric"), None);
    assert_eq!(column_to_metric(&LeaderboardKind::OpenLLMv2, ""), None);
}

// ═══════════════════════════════════════════════════════════════════════
// CLAIM: Custom leaderboard kind accepts arbitrary dataset IDs
// ═══════════════════════════════════════════════════════════════════════

#[test]
fn falsify_custom_leaderboard_passthrough() {
    let kind = LeaderboardKind::Custom("my-org/my-leaderboard".into());
    assert_eq!(kind.dataset_repo_id(), "my-org/my-leaderboard");
    // Custom defaults to Accuracy
    assert_eq!(kind.primary_metric(), Metric::Accuracy);
}

#[test]
fn falsify_custom_leaderboard_generic_mapping() {
    let kind = LeaderboardKind::Custom("anything".into());
    // Generic mapper should handle common metric names
    assert_eq!(column_to_metric(&kind, "accuracy"), Some(Metric::Accuracy));
    assert_eq!(column_to_metric(&kind, "wer"), Some(Metric::WER));
    assert_eq!(column_to_metric(&kind, "bleu"), Some(Metric::BLEU));
    assert_eq!(column_to_metric(&kind, "rouge1"), Some(Metric::ROUGE(RougeVariant::Rouge1)));
    assert_eq!(column_to_metric(&kind, "rougel"), Some(Metric::ROUGE(RougeVariant::RougeL)));
    assert_eq!(column_to_metric(&kind, "perplexity"), Some(Metric::Perplexity));
    assert_eq!(column_to_metric(&kind, "ppl"), Some(Metric::Perplexity));
    assert_eq!(column_to_metric(&kind, "mmlu"), Some(Metric::MMLUAccuracy));
    assert_eq!(column_to_metric(&kind, "pass@1"), Some(Metric::PassAtK(1)));
    assert_eq!(column_to_metric(&kind, "ndcg@10"), Some(Metric::NDCGAtK(10)));
}

// ═══════════════════════════════════════════════════════════════════════
// CLAIM: HfLeaderboard.find_model returns correct entry
// ═══════════════════════════════════════════════════════════════════════

#[test]
fn falsify_find_model_exact_match() {
    let mut lb = HfLeaderboard::new(LeaderboardKind::OpenASR);
    let mut e = LeaderboardEntry::new("openai/whisper-large-v3");
    e.scores.insert("wer".into(), 0.05);
    lb.entries.push(e);
    lb.entries.push(LeaderboardEntry::new("other/model"));

    let found = lb.find_model("openai/whisper-large-v3");
    assert!(found.is_some());
    assert_eq!(found.expect("operation should succeed").get_score("wer"), Some(0.05));
}

#[test]
fn falsify_find_model_no_partial_match() {
    let mut lb = HfLeaderboard::new(LeaderboardKind::OpenASR);
    lb.entries.push(LeaderboardEntry::new("openai/whisper-large-v3"));

    // Partial match should NOT work
    assert!(lb.find_model("openai/whisper").is_none());
    assert!(lb.find_model("whisper-large-v3").is_none());
}

// ═══════════════════════════════════════════════════════════════════════
// CLAIM: Empty leaderboard produces empty Leaderboard
// ═══════════════════════════════════════════════════════════════════════

#[test]
fn falsify_empty_hf_leaderboard() {
    let hf = HfLeaderboard::new(LeaderboardKind::MTEB);
    let lb = to_leaderboard(&hf);
    assert!(lb.results.is_empty());
    assert!(lb.best().is_none());
}

// ═══════════════════════════════════════════════════════════════════════
// CLAIM: Scores with no matching metric columns are silently dropped
// ═══════════════════════════════════════════════════════════════════════

#[test]
fn falsify_unrecognized_columns_dropped() {
    let mut hf = HfLeaderboard::new(LeaderboardKind::OpenASR);
    let mut entry = LeaderboardEntry::new("model-a");
    entry.scores.insert("wer".into(), 0.1);
    entry.scores.insert("completely_unknown_column".into(), 42.0);
    entry.scores.insert("another_unknown".into(), 99.0);
    hf.entries.push(entry);

    let lb = to_leaderboard(&hf);
    let result = &lb.results[0];
    // Only WER should be present
    assert_eq!(result.get_score(Metric::WER), Some(0.1));
    // Unknown columns should not create spurious metrics
    assert_eq!(result.scores.len(), 1);
}