Sarthak213/doosra-toolcalls

Doosra tool calls

8,000 conversations in which a cricket question is answered by calling Doosra's analytics tools and then writing an answer from their results. It was built to fine-tune a small model (Qwen3.5 4B) to use Doosra's 25 tools as reliably as a larger one.

Split	Conversations
train	6,695
validation	345
test	960

The test split uses held-out entities: about 15% of players, teams and venues appear only there. The test measures generalisation to names the model never saw.

How it was made

Nothing here was written by hand, and nothing was generated by an LLM without checking.

1. Questions from templates over real entities. Intents cover every tool: leaderboards, player splits, comparisons, form, match-ups, venues, records, luck (FIBS), match replays, data coverage, and multi-step requests (a table then a chart; the same player in two scopes). They use real players, teams, venues, competitions and seasons from the Doosra database, picked in proportion to how much they have played.
2. Gold tool calls, executed. Each question's correct calls are known by construction and were run against the database exactly as the Doosra agent runs them. Calls that errored or returned nothing were dropped, so every conversation is grounded in real data. Tool results appear as the model sees them in the app (JSON, tables capped at 25 rows).
3. Answers from the results only. The final final_answer call is written from the tool results in Doosra's style: the direct answer first, bold names and key numbers, the scope used, and caveats from the results' notes. Every number in an answer appears in a tool result.
4. Paraphrased. About 46% of questions were rewritten by Qwen3.5 9B, running locally, to vary the wording. A rewrite was kept only if every name, competition and number survived, so the gold calls still match.
5. Checked. Every answer quotes figures, and only figures that appear in its tool results; no conversation copies or answers a question from Doosra's end-to-end evaluation (its players, grounds and facts are left out).

Version 2

Version 1 trained models that aced questions shaped like its templates and did worse than the base model on open questions. Evaluating them showed why, and version 2 fixes it:

• Answers that said nothing. Team records, data-coverage and some split answers had no figures (a template bug); records showed empty scores. All fixed, and answers now lead with the figure the question asks for.
• Single figures. "How many sixes / balls faced / dismissals / maidens…" questions, which need player_stats's metrics argument: version 1 had almost none.
• SQL. About 20 query patterns on the real schema (version 1 had 4 conversations); the compact tool list now keeps run_sql's schema.
• Recovering. Conversations where a call goes wrong and the next fixes it: a SQL error and the corrected query; a result without the asked figure and a second call with metrics.
• Venue records, men's/women's wording, head-to-heads.

Format

Each line holds id, intent, question (and original_question if paraphrased), and messages in OpenAI chat format:

• system: the short prompt;
• user: the question;
• an assistant tool call and its tool result, repeated for each call;
• a final assistant call to final_answer.

The tool definitions every example uses are in tools.json.

Intents

Intent	Conversations
leaderboard	1,592
stat_one	594
stats_split	583
profile	504
compare	501
stats_scope	456
form	355
similar	268
venue	259
chart	255
percentiles	238
luck	225
two_scopes	218
records	193
entry	186
lookup	162
arc	160
venue_record	158
matchup	151
open	149
team_record	124
stat_recover	117
sql	114
replay	110
team_board	95
compare_chart	72
matrix	56
coverage	41
sql_recover	38
fibs	21
search_metrics	5

Licence

The cricket data behind the tool results is from Cricsheet, under the Open Data Commons Attribution License (ODC-By 1.0). This dataset is released under the same licence.