Meet8-0.6B
An experimental full-parameter post-training run by Ma7ee7, starting from Qwen/Qwen3-0.6B-Base.
Training
- Supervised fine-tuning on instruction, identity, human-written chat, generated FineWeb instruction examples, and generated creative writing.
- Paired direct-answer and
/thinkwarmup using short reasoning examples. - Full-parameter GRPO on math, science, and logic tasks.
- Approximately 3,500 total RL optimizer steps.
Usage
Use the saved tokenizer chat template. The following system message substantially improves identity and reasoning-mode compliance:
You are Meet8-0.6B, developed by Ma7ee7. If the user's message begins with /think, give a brief explanation inside ..., then give your final response outside the tags. Otherwise answer directly without thinking tags. Never return an empty answer.
Diagnostic results
Small held-out greedy evaluations, with 32 questions per task and mode:
| Task | Direct accuracy | Think accuracy |
|---|---|---|
| GSM8K | 12.50% | 37.50% |
| ARC-Easy | 81.25% | 84.38% |
| Arithmetic | 56.25% | 84.38% |
| Procedural logic | 56.25% | 50.00% |
| RuleTaker | 53.13% | 46.88% |
| Knights and knaves | 28.13% | 15.63% |
| Countdown | 3.13% | 0.00% |
These are small custom diagnostics, not standard full-split benchmark scores. Direct and think partitions may contain different questions. No successful standardized Base-versus-Meet8 comparison is available.
Limitations
Creative writing can enter severe repetition loops.
Reasoning explanations can be incorrect despite appearing plausible.
Identity and /think behavior depend strongly on the system prompt.
Overall improvement over the base model has not been established.
Data and licensing
Training included HuggingFaceH4/no_robots, distributed under CC BY-NC 4.0. Consult upstream model and dataset licenses before reuse. Additional logic datasets and generated data were used; this card does not provide a complete provenance or overlap audit.