Spaces:
Running
on
CPU Upgrade
Running
on
CPU Upgrade
update api routes
Browse files
backend/data/lighteval_results/lighteval_results.json
CHANGED
@@ -23,8 +23,8 @@
|
|
23 |
{
|
24 |
"model": "deepseek-ai/DeepSeek-R1-Distill-Llama-70B",
|
25 |
"provider": "sambanova",
|
26 |
-
"accuracy":
|
27 |
-
"execution_time":
|
28 |
-
"status": "
|
29 |
}
|
30 |
]
|
|
|
23 |
{
|
24 |
"model": "deepseek-ai/DeepSeek-R1-Distill-Llama-70B",
|
25 |
"provider": "sambanova",
|
26 |
+
"accuracy": 1.0,
|
27 |
+
"execution_time": 27.880253076553345,
|
28 |
+
"status": "success"
|
29 |
}
|
30 |
]
|
backend/data/lighteval_results/results/deepseek-ai/DeepSeek-R1-Distill-Llama-70B/results_2025-03-28T12-59-46.530720.json
ADDED
@@ -0,0 +1,121 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"config_general": {
|
3 |
+
"lighteval_sha": "?",
|
4 |
+
"num_fewshot_seeds": 1,
|
5 |
+
"override_batch_size": null,
|
6 |
+
"max_samples": 30,
|
7 |
+
"job_id": 0,
|
8 |
+
"start_time": 193754.29830825,
|
9 |
+
"end_time": 193775.660671041,
|
10 |
+
"total_evaluation_time_secondes": "21.362362790998304",
|
11 |
+
"model_name": "deepseek-ai/DeepSeek-R1-Distill-Llama-70B",
|
12 |
+
"model_sha": "",
|
13 |
+
"model_dtype": null,
|
14 |
+
"model_size": "",
|
15 |
+
"generation_parameters": {
|
16 |
+
"early_stopping": null,
|
17 |
+
"repetition_penalty": null,
|
18 |
+
"frequency_penalty": null,
|
19 |
+
"length_penalty": null,
|
20 |
+
"presence_penalty": null,
|
21 |
+
"max_new_tokens": null,
|
22 |
+
"min_new_tokens": null,
|
23 |
+
"seed": null,
|
24 |
+
"stop_tokens": null,
|
25 |
+
"temperature": null,
|
26 |
+
"top_k": null,
|
27 |
+
"min_p": null,
|
28 |
+
"top_p": null,
|
29 |
+
"truncate_prompt": null,
|
30 |
+
"response_format": null
|
31 |
+
}
|
32 |
+
},
|
33 |
+
"results": {
|
34 |
+
"custom|yourbench|0": {
|
35 |
+
"accuracy": 1.0,
|
36 |
+
"accuracy_stderr": 0.0
|
37 |
+
},
|
38 |
+
"all": {
|
39 |
+
"accuracy": 1.0,
|
40 |
+
"accuracy_stderr": 0.0
|
41 |
+
}
|
42 |
+
},
|
43 |
+
"versions": {
|
44 |
+
"custom|yourbench|0": 0
|
45 |
+
},
|
46 |
+
"config_tasks": {
|
47 |
+
"custom|yourbench": {
|
48 |
+
"name": "yourbench",
|
49 |
+
"prompt_function": "yourbench_prompt",
|
50 |
+
"hf_repo": "yourbench/yourbench_d0766aeb-d261-4f0f-870c-537432fd8584",
|
51 |
+
"hf_subset": "multi_hop_questions",
|
52 |
+
"metric": [
|
53 |
+
{
|
54 |
+
"metric_name": [
|
55 |
+
"accuracy"
|
56 |
+
],
|
57 |
+
"higher_is_better": {
|
58 |
+
"accuracy": true
|
59 |
+
},
|
60 |
+
"category": "7",
|
61 |
+
"use_case": "1",
|
62 |
+
"sample_level_fn": "compute",
|
63 |
+
"corpus_level_fn": {
|
64 |
+
"accuracy": "mean"
|
65 |
+
}
|
66 |
+
}
|
67 |
+
],
|
68 |
+
"hf_revision": null,
|
69 |
+
"hf_filter": null,
|
70 |
+
"hf_avail_splits": [
|
71 |
+
"train"
|
72 |
+
],
|
73 |
+
"trust_dataset": true,
|
74 |
+
"evaluation_splits": [
|
75 |
+
"train"
|
76 |
+
],
|
77 |
+
"few_shots_split": null,
|
78 |
+
"few_shots_select": null,
|
79 |
+
"generation_size": 8192,
|
80 |
+
"generation_grammar": null,
|
81 |
+
"stop_sequence": [],
|
82 |
+
"num_samples": null,
|
83 |
+
"suite": [
|
84 |
+
"custom"
|
85 |
+
],
|
86 |
+
"original_num_docs": 34,
|
87 |
+
"effective_num_docs": 30,
|
88 |
+
"must_remove_duplicate_docs": false,
|
89 |
+
"version": 0
|
90 |
+
}
|
91 |
+
},
|
92 |
+
"summary_tasks": {
|
93 |
+
"custom|yourbench|0": {
|
94 |
+
"hashes": {
|
95 |
+
"hash_examples": "8deb6ee598efe642",
|
96 |
+
"hash_full_prompts": "ee276216c7fba0dc",
|
97 |
+
"hash_input_tokens": "79ab129e9a18c6d6",
|
98 |
+
"hash_cont_tokens": "79ab129e9a18c6d6"
|
99 |
+
},
|
100 |
+
"truncated": 0,
|
101 |
+
"non_truncated": 30,
|
102 |
+
"padded": 0,
|
103 |
+
"non_padded": 30,
|
104 |
+
"effective_few_shots": 0.0,
|
105 |
+
"num_truncated_few_shots": 0
|
106 |
+
}
|
107 |
+
},
|
108 |
+
"summary_general": {
|
109 |
+
"hashes": {
|
110 |
+
"hash_examples": "134194cd9d247350",
|
111 |
+
"hash_full_prompts": "59b03121730720e8",
|
112 |
+
"hash_input_tokens": "05a66e44e190c178",
|
113 |
+
"hash_cont_tokens": "05a66e44e190c178"
|
114 |
+
},
|
115 |
+
"truncated": 0,
|
116 |
+
"non_truncated": 30,
|
117 |
+
"padded": 0,
|
118 |
+
"non_padded": 30,
|
119 |
+
"num_truncated_few_shots": 0
|
120 |
+
}
|
121 |
+
}
|
frontend/server.js
CHANGED
@@ -30,22 +30,7 @@ app.use((req, res, next) => {
|
|
30 |
next();
|
31 |
});
|
32 |
|
33 |
-
//
|
34 |
-
app.use(
|
35 |
-
createProxyMiddleware({
|
36 |
-
target: backendUrl,
|
37 |
-
changeOrigin: true,
|
38 |
-
pathRewrite: {
|
39 |
-
"^/api": "", // Remove /api prefix if present
|
40 |
-
},
|
41 |
-
onError: (err, req, res) => {
|
42 |
-
console.error("Proxy Error:", err);
|
43 |
-
res.status(500).json({ error: "Proxy Error", details: err.message });
|
44 |
-
},
|
45 |
-
})
|
46 |
-
);
|
47 |
-
|
48 |
-
// Serve static files from the build directory
|
49 |
app.use(
|
50 |
express.static(path.join(__dirname, "build"), {
|
51 |
// Don't cache HTML files
|
@@ -62,6 +47,33 @@ app.use(
|
|
62 |
})
|
63 |
);
|
64 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
65 |
// Handle all other routes by serving index.html
|
66 |
app.get("*", (req, res) => {
|
67 |
// Headers for client-side routing
|
|
|
30 |
next();
|
31 |
});
|
32 |
|
33 |
+
// Serve static files from the build directory first
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
34 |
app.use(
|
35 |
express.static(path.join(__dirname, "build"), {
|
36 |
// Don't cache HTML files
|
|
|
47 |
})
|
48 |
);
|
49 |
|
50 |
+
// Proxy only API routes to the Python backend
|
51 |
+
app.use(
|
52 |
+
[
|
53 |
+
"/health",
|
54 |
+
"/upload",
|
55 |
+
"/generate-benchmark",
|
56 |
+
"/config-logs",
|
57 |
+
"/benchmark-logs",
|
58 |
+
"/benchmark-questions",
|
59 |
+
"/evaluate-benchmark",
|
60 |
+
"/evaluation-logs",
|
61 |
+
"/evaluation-results",
|
62 |
+
"/download-dataset",
|
63 |
+
],
|
64 |
+
createProxyMiddleware({
|
65 |
+
target: backendUrl,
|
66 |
+
changeOrigin: true,
|
67 |
+
pathRewrite: {
|
68 |
+
"^/api": "", // Remove /api prefix if present
|
69 |
+
},
|
70 |
+
onError: (err, req, res) => {
|
71 |
+
console.error("Proxy Error:", err);
|
72 |
+
res.status(500).json({ error: "Proxy Error", details: err.message });
|
73 |
+
},
|
74 |
+
})
|
75 |
+
);
|
76 |
+
|
77 |
// Handle all other routes by serving index.html
|
78 |
app.get("*", (req, res) => {
|
79 |
// Headers for client-side routing
|