Fix HF dataset import internal error and training SQLite readiness.

Wrap AssistentImportHfDataset in try/catch, batch inserts in a transaction, clearer HF rows errors for gated datasets, EnsureTrainingReady with actionable message, and friendlier training UI errors instead of generic internal error.

Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
Leonid Pershin
2026-08-23 20:11:18 +03:00
co-authored by Cursor
parent b0f2736f65
commit 63538fe738
9 changed files with 323 additions and 155 deletions
+121 -72
View File
@@ -426,94 +426,143 @@ public partial class SwarmAssistentExtension
public async Task<JObject> AssistentImportHfDataset(Session session, JObject raw)
{
string dataset = raw?["dataset"]?.ToString();
int limit = raw?["limit"]?.Value<int?>() ?? 200;
JObject mapping = raw?["mapping"] as JObject;
string id = NormalizeHfDatasetId(dataset);
if (id is null)
try
{
return new JObject { ["error"] = "invalid dataset id" };
}
JObject check = await CheckHfDatasetInternal(session, id, useCache: true);
string gate = check["gate"]?.ToString();
if (gate == "rejected")
{
return new JObject { ["error"] = check["reason"]?.ToString() ?? "rejected" };
}
if (MappingRequired(check, mapping))
{
return new JObject { ["error"] = "Нужен маппинг колонок", ["check"] = check };
}
mapping = ResolveHfMapping(check, mapping);
int take = Math.Clamp(limit, 1, 5000);
JArray rows = [];
string config = check["config"]?.ToString() ?? "default";
string split = check["split"]?.ToString() ?? "train";
int offset = 0;
while (rows.Count < take)
{
int chunk = Math.Min(100, take - rows.Count);
using HttpRequestMessage rowsReq = HfRequest($"{HfDatasetsServer}/rows?dataset={Uri.EscapeDataString(id)}&config={Uri.EscapeDataString(config)}&split={Uri.EscapeDataString(split)}&offset={offset}&length={chunk}", session);
using HttpResponseMessage rowsResp = await HttpClient.SendAsync(rowsReq);
string rowsBody = await rowsResp.Content.ReadAsStringAsync();
if (!rowsResp.IsSuccessStatusCode)
string dataset = raw?["dataset"]?.ToString();
int limit = raw?["limit"]?.Value<int?>() ?? 200;
JObject mapping = raw?["mapping"] as JObject;
string id = NormalizeHfDatasetId(dataset);
if (id is null)
{
break;
return new JObject { ["error"] = "invalid dataset id" };
}
JObject parsed = JObject.Parse(rowsBody);
JArray batch = parsed["rows"] as JArray ?? [];
if (batch.Count == 0)
JObject check = await CheckHfDatasetInternal(session, id, useCache: true);
string gate = check["gate"]?.ToString();
if (gate == "rejected")
{
break;
return new JObject { ["error"] = check["reason"]?.ToString() ?? "rejected" };
}
foreach (JToken t in batch)
if (MappingRequired(check, mapping))
{
rows.Add(t);
return new JObject { ["error"] = "Нужен маппинг колонок", ["check"] = check };
}
offset += batch.Count;
if (batch.Count < chunk)
mapping = ResolveHfMapping(check, mapping);
int take = Math.Clamp(limit, 1, 5000);
JArray rows = [];
string config = check["config"]?.ToString() ?? "default";
string split = check["split"]?.ToString() ?? "train";
int offset = 0;
while (rows.Count < take)
{
break;
int chunk = Math.Min(100, take - rows.Count);
using HttpRequestMessage rowsReq = HfRequest($"{HfDatasetsServer}/rows?dataset={Uri.EscapeDataString(id)}&config={Uri.EscapeDataString(config)}&split={Uri.EscapeDataString(split)}&offset={offset}&length={chunk}", session);
using HttpResponseMessage rowsResp = await HttpClient.SendAsync(rowsReq);
string rowsBody = await rowsResp.Content.ReadAsStringAsync();
if (!rowsResp.IsSuccessStatusCode)
{
if (rows.Count == 0)
{
return new JObject
{
["error"] = $"HF rows HTTP {(int)rowsResp.StatusCode}: {Clip(rowsBody, 240)}. "
+ "Для gated/NSFW добавь huggingface_api в User Settings.",
};
}
break;
}
JObject parsed = JObject.Parse(rowsBody);
JArray batch = parsed["rows"] as JArray ?? [];
if (batch.Count == 0)
{
break;
}
foreach (JToken t in batch)
{
rows.Add(t);
}
offset += batch.Count;
if (batch.Count < chunk)
{
break;
}
}
}
if (rows.Count == 0)
{
rows = check["sample_rows"] as JArray ?? [];
}
int imported = 0;
foreach (JToken rowTok in rows.Take(take))
{
if (rowTok is not JObject row)
if (rows.Count == 0)
{
continue;
rows = check["sample_rows"] as JArray ?? [];
}
JObject rowData = row["row"] as JObject ?? row;
JArray messages = ConvertHfRowToMessages(rowData, check["schema"] as JObject, mapping);
if (messages is null || messages.Count == 0)
List<JObject> toSave = [];
foreach (JToken rowTok in rows.Take(take))
{
continue;
if (rowTok is not JObject row)
{
continue;
}
JObject rowData = row["row"] as JObject ?? row;
JArray messages = ConvertHfRowToMessages(rowData, check["schema"] as JObject, mapping);
if (messages is null || messages.Count == 0)
{
continue;
}
toSave.Add(new JObject
{
["source"] = "hf",
["hf_repo"] = id,
["messages"] = messages,
["status"] = "draft",
});
}
Memory.UpsertTrainSample(new JObject
int imported = toSave.Count > 0 ? Memory.ImportTrainSamplesBatch(toSave) : 0;
if (imported == 0 && check["runner_only"]?.Value<bool?>() == true)
{
["source"] = "hf",
["hf_repo"] = id,
["messages"] = messages,
return new JObject { ["success"] = true, ["imported"] = 0, ["runner_only"] = true, ["id"] = id, ["note"] = "Большой набор — используй HF id в QLoRA-раннере" };
}
if (imported == 0)
{
return new JObject
{
["error"] = rows.Count == 0
? "HF не отдал строки — проверь token (gated/NSFW) и маппинг колонок"
: "0 строк после маппинга — проверь колонки user/assistant",
["rows_fetched"] = rows.Count,
["id"] = id,
};
}
return new JObject
{
["success"] = true,
["imported"] = imported,
["id"] = id,
["rows_fetched"] = rows.Count,
["status"] = "draft",
});
imported++;
};
}
if (imported == 0 && check["runner_only"]?.Value<bool?>() == true)
catch (Exception ex)
{
return new JObject { ["success"] = true, ["imported"] = 0, ["runner_only"] = true, ["id"] = id, ["note"] = "Большой набор — используй HF id в QLoRA-раннере" };
Logs.Error($"AssistentImportHfDataset: {ex}");
return new JObject { ["error"] = ex.Message };
}
return new JObject
}
static string HfCellString(JToken tok)
{
if (tok is null || tok.Type == JTokenType.Null)
{
["success"] = true,
["imported"] = imported,
["id"] = id,
["rows_fetched"] = rows.Count,
["status"] = "draft",
};
return "";
}
if (tok is JArray arr)
{
List<string> parts = [];
foreach (JToken t in arr)
{
string s = t?.Type == JTokenType.String ? t.ToString() : t?.ToString(Newtonsoft.Json.Formatting.None);
if (!string.IsNullOrWhiteSpace(s))
{
parts.Add(s.Trim());
}
}
return string.Join(", ", parts);
}
return tok.ToString().Trim();
}
static JArray ConvertHfRowToMessages(JObject row, JObject schema, JObject mapping)
@@ -570,14 +619,14 @@ public partial class SwarmAssistentExtension
}
if (kind == "fiction_tags_text")
{
string text = row["text"]?.ToString() ?? "";
string text = HfCellString(row["text"]);
if (string.IsNullOrWhiteSpace(text))
{
return null;
}
List<string> userParts = [];
string title = row["title"]?.ToString()?.Trim();
string tags = row["tags"]?.ToString()?.Trim();
string title = HfCellString(row["title"]);
string tags = HfCellString(row["tags"]);
if (!string.IsNullOrWhiteSpace(title))
{
userParts.Add($"Title: {title}");