# The sharegpt set

Does `funes recall` find a conversation when it is in Chinese? This set asks 30 questions, each in
Chinese and in English, of 2000 programming conversations held in both languages, and scores where
the conversation the question is about ranks among recall's hits. It was built for
[funes#144](https://github.com/huggingface/funes/issues/144) (multilingual embedding models).

## Memories

| memory | language | embedding model |
|---|---|---|
| [`huggingface/funes-sharegpt-zh`](https://huggingface.co/datasets/huggingface/funes-sharegpt-zh) | Chinese | `BAAI/bge-small-en-v1.5` |
| [`huggingface/funes-sharegpt-en`](https://huggingface.co/datasets/huggingface/funes-sharegpt-en) | English | `BAAI/bge-small-en-v1.5` |
| [`huggingface/funes-sharegpt-zh-e5`](https://huggingface.co/datasets/huggingface/funes-sharegpt-zh-e5) | Chinese | `intfloat/multilingual-e5-small` |

The three hold the same conversations. `funes-sharegpt-zh-e5` needs funes 1.8.0 or later, and its
text index was rebuilt with funes 1.8.0's `icu` analyzer on 2026-10-08.

## Corpus

2000 "Program and Code" conversations from
[`shareAI/ShareGPT-Chinese-English-90k`](https://huggingface.co/datasets/shareAI/ShareGPT-Chinese-English-90k)
(Apache-2.0), `computer_zh_26k(fixed).jsonl` and `computer_en_26k(fixed).jsonl`, paired by
`conversation_id`. The sample (seed 144) is drawn from the 3833 conversations present in both files
with the same number of turns. [tools/convert.py](tools/convert.py) writes each conversation as one
`.funes.jsonl` session, one turn per message, harness `sharegpt`, and `funes index` reads the
directory. The conversations are chats with a model, not agent sessions: they hold no tool calls.

## `questions-en.jsonl`, `questions-zh.jsonl`

One question a line, the same 30 in each file, in that file's language:

```json
{"id": "s01", "question": "…", "gold": [{"session_id": "<conversation id>"}], "quote": "…"}
```

`gold` names the conversation, any turn of it.

The 30 conversations were drawn at random from the corpus (seed 30). Each English question says what
the conversation was after in other words than its first message. The Chinese question translates
it the way a Chinese developer would type it, product names (React, Laravel, Pinecone) left in
English. Three Chinese questions hold no Latin letters at all (s12, s27, s28). `quote` is the first
300 characters of the conversation's first English message, kept for review only. Claude wrote the
questions on 2026-10-06.

## Scoring

A hit counts when it comes from the gold conversation. Each file is asked once, at `funes recall`'s
defaults, of a memory in its language, and the receipt reports `found@8` (the gold among the eight
hits a default recall returns), `hit@1`, `hit@5` and the MRR of the first such hit.

## Run

`score_recall` (`benchmark/score_recall.rs` in the funes repository) asks the questions and prints
the receipt. From a funes checkout, with this dataset at `<bench>`:

```bash
cargo run --release --example score_recall -- <bench>/sharegpt/questions-zh.jsonl --memory huggingface/funes-sharegpt-zh-e5
```

## Results

Run of 2026-10-13 with `score_recall`, funes at main `44906fad`, on the 30 questions of each file. Full
output and per-question ranks: [results/2026-10-13.md](results/2026-10-13.md).

| memory | questions | hit@1 | hit@5 | found@8 | MRR | p50 |
|---|---|---|---|---|---|---|
| `funes-sharegpt-en` | English | 28 | 30 | 30 | 0.961 | 1.07 s |
| `funes-sharegpt-zh` | Chinese | 16 | 23 | 23 | 0.636 | 1.05 s |
| `funes-sharegpt-zh-e5` | Chinese | 28 | 30 | 30 | 0.957 | 1.05 s |

Run of 2026-10-06, funes at commit `fe60450`. Full receipt and per-question ranks:
[results/2026-10-06.md](results/2026-10-06.md). One question is 3.3 points of 30. A script
`score_recall` replaced scored this run: it ranked the sessions the hits name, in three recalls per
question (the default, the fused order 30 deep, and a pool of 40 reranked), which the columns keep.

| arm | default found@8 | default MRR | rerank found@8 | rerank MRR |
|---|---|---|---|---|
| 1. English, `bge-small-en-v1.5`, default analyzer | 30 | 0.967 | 30 | 0.923 |
| 2. Chinese, `bge-small-en-v1.5`, default analyzer (funes 1.7) | 23 | 0.601 | 28 | 0.917 |
| 3. Chinese, `multilingual-e5-small`, default analyzer | 30 | 0.867 | 30 | 0.983 |
| 4. Chinese, `bge-small-en-v1.5`, `icu` analyzer | 30 | 0.865 | 30 | 0.983 |
| 5. Chinese, `multilingual-e5-small`, `icu` analyzer | 30 | 0.983 | 30 | 0.983 |

The default analyzer splits text only on spaces and punctuation, so a Chinese clause is one token,
and drops tokens of 40 bytes or more (from 14 Chinese characters on). `icu` segments Chinese into words.
Arms 3 to 5 ran a prototype of the changes released in funes 1.8.0, described in the receipt.

Limits of this run: 30 questions written by one author, a corpus of 7.6k chunks where a real memory
holds 100k or more, chats rather than agent sessions, and the Chinese side of the source dataset may
be machine-translated.