A model can return the correct root while giving a flawed explanation. In FU-max-boop’s public Spark-X2.5-1.7B case, the saved response to sqrt(x) = x - 2 correctly ends with FINAL: [4], but calls x >= 2 the expression’s domain.
The expression domain is x >= 0. Every solution must additionally satisfy x >= 2 because the square root is nonnegative. At x = 1, the expressions are defined but unequal: 1 ≠ −1. At x = 4, both sides equal 2. The response’s root and substitution checks remain valid, while its explanation of the domain needs correction.
Watch “Right answer, wrong explanation?” by SparkLLM
The participant reports a partial local Q4_K_M study: 96 requests planned, 45 completed, one interrupted for resource limits and 50 unattempted. Its strict final-answer score is 21/45. This is a result on that partial diagnostic with a required answer format, not general model accuracy. The report includes prompts, raw responses, scoring code and limitations. This article explains the participant’s public evidence; it does not claim a new model inference run.
Read the original case and inspect the pinned report and evidence.
Bring your own real, reproducible math evaluation to HER Hack-Astron #6. Here are the current answers to 11 participation questions, updated September 11, 2026.
1. Are there gender restrictions for individual entrants?
No. Individual entrants may participate regardless of gender.
2. Who does the at-least-50% women-contributors rule apply to?
Team entries only: at least 50% of a team’s contributors must be women. This percentage requirement does not apply to individual entrants.
3. Can AI assist with the work or execute experiments?
Yes. Both AI assistance and AI execution are allowed; disclose honestly which parts involved AI.
4. Can an AI persona be listed as a real human contributor?
No. Do not present an AI persona as a real human contributor or count one toward the team’s human-contributor percentages.
5. What is the basic requirement for the evaluation?
It must be real and reproducible. Retain evidence of the actual run and results, such as the model revision, problems and prompts, runtime settings, raw outputs and scoring method. Report failures and limitations honestly.
6. How do I complete a formal submission?
Publish the complete evaluation in the Hugging Face Discussions for the model you used, with HER Hack-Astron #6 in the title. Then reply to event GitHub Issue #9 with the direct Discussion link. Complete both steps for a formal submission.
7. How can international winners receive a prize?
International winners may use an available payout method, such as PayPal or bank transfer.
8. Can winners in mainland China receive RMB through a domestic method?
An RMB payout through an available domestic method can be discussed. The specific arrangement depends on what is available and confirmed by both sides.
9. Is an upfront fee required to enter or claim a prize?
No. There is no upfront participation or prize-claim fee.
10. When should identity or payment information be provided?
Provide necessary sensitive identity or payment information privately only after winning, and only when needed. Do not include it in public Discussions or Issues.
11. Is a live presentation currently required?
No live presentation is currently required. Submit the required real, reproducible evaluation materials.
Ready to enter?
Run a real math evaluation with Spark-X2.5-1.7B or 4B, publish the complete results and limitations, and follow the formal submission steps. Deadline: September 13, 2026, 24:00 Beijing time (UTC+8). There is one winning prize of USD 100; the payout method is subject to the available arrangements described above.
The formal submission has two parts: publish the complete result in the Hugging Face Discussions for the model you used, with HER Hack-Astron #6 in the title, and reply to the event’s GitHub Issue #9 with the direct link.
Full event rules and submission · Spark-X2.5 project
SparkX25 #HERHackAstron #MathReasoning #ModelEvaluation #OpenSourceAI #Reproducibility
Video music: El Rizo de Oro, Op.16 by Trinidad Moreno; synthesized piano by Eniel García. IMSLP, CC0 1.0.
Top comments (0)