Add inference samples at step 9,131,773 (698M, 7.19B/35.7B tokens)

samples/inference_samples_step09131773.md

@@ -0,0 +1,145 @@
+# AGILLM-3-large Inference Samples
+
+**Model:** AGILLM-3-large  
+**Parameters:** 698M  
+**Checkpoint:** pretrain_step09131773  
+**Training Progress:** 7.19B / 35.7B tokens  
+**Generated:** 2026-02-18 19:20:11 UTC  
+**Hardware:** RTX 3060 12GB (inference)  
+
+---
+
+## AR Mode (Autoregressive)
+
+Standard left-to-right token generation.
+
+### Prompt: "The meaning of life is"
+
+> The meaning of life is a necessary part of the project. And it has become much more difficult to reach these creatures – that's when a pet's journey began.
+
+*[3.23s | 200 tokens | 62.0 tok/s]*
+
+---
+
+### Prompt: "In the year 2050, artificial intelligence"
+
+> In the year 2050, artificial intelligence (AI) has reached half a billion people . This is despite the incredible progress that AI’s last year.
+
+*[3.17s | 200 tokens | 63.2 tok/s]*
+
+---
+
+### Prompt: "Once upon a time, a young scientist discovered"
+
+> Once upon a time, a young scientist discovered it in 1997 and he was working on a new game: the arc of the social ladder.
+
+*[3.24s | 200 tokens | 61.7 tok/s]*
+
+---
+
+### Prompt: "The stock market crashed today because"
+
+> The stock market crashed today because of the current share price war, as well as a recession in the U.S. — and also with more than 20% market share.
+
+*[3.07s | 200 tokens | 65.2 tok/s]*
+
+---
+
+### Prompt: "Dear diary, today I learned that"
+
+> Dear diary, today I learned that 16 weeks of work and school staff were better off supporting students who got the course to leave their job. The research team also evaluated reported that just over half of students have had already used the most%, not only for those who spend less than two years in the tech sector, but even now around 20% of students they didn’t know how to stop working or quit.
+
+*[3.02s | 200 tokens | 66.1 tok/s]*
+
+---
+
+### Prompt: "Breaking news: scientists have confirmed that"
+
+> Breaking news: scientists have confirmed that some artificial intelligence models, such as OpenAI’s GPT-4o, won’t answer questions about them. But we still don’t know what will happen to the actor.
+
+*[3.06s | 200 tokens | 65.4 tok/s]*
+
+---
+
+### Prompt: "The recipe for happiness includes"
+
+> The recipe for happiness includes a drink of Ariana Grande and Jeff Richards, the coelist to Shank Geim.
+
+*[3.01s | 200 tokens | 66.5 tok/s]*
+
+---
+
+### Prompt: "Python is a programming language that"
+
+> Python is a programming language that makes it much easier to be a regular item. At the end of the process, you have 3rd best practices for running an animated painting.
+
+*[3.02s | 200 tokens | 66.3 tok/s]*
+
+---
+
+### Prompt: "The history of the Roman Empire begins with"
+
+> The history of the Roman Empire begins with a handful of sparsely used footage. The footage, which was released last Tuesday and showed an ambitious vision for regenerating today’s antics from September 1995 to November 2016.
+
+*[3.04s | 200 tokens | 65.8 tok/s]*
+
+---
+
+### Prompt: "If I could travel anywhere in the world"
+
+> If I could travel anywhere in the world to see work and for this reason it would be a good deal," she said.
+
+*[3.03s | 200 tokens | 66.0 tok/s]*
+
+---
+
+## SAT Mode (Speculative Autoregressive Tokens)
+
+Multi-token parallel generation via the SAT head. Faster but less coherent at this training stage.
+
+### Prompt: "The meaning of life is"
+
+> The meaning of life is de la. losía del are tu.
+
+*[1.96s | 200 tokens | 102.2 tok/s]*
+
+---
+
+### Prompt: "In the year 2050, artificial intelligence"
+
+> In the year 2050, artificial intelligence a de thatar laña losÍes tu no son aquí uno at which with: afectos suía con. compart theiró un ha.
+
+*[1.97s | 200 tokens | 101.5 tok/s]*
+
+---
+
+### Prompt: "Once upon a time, a young scientist discovered"
+
+> Once upon a time, a young scientist discovered la de thatá en sus del.ando lasÍ con aquí los su tu aciónña compartar:os cambadosadaia teir.
+
+*[1.96s | 200 tokens | 102.0 tok/s]*
+
+---
+
+### Prompt: "Python is a programming language that"
+
+> Python is a programming language that la.
+
+*[1.96s | 200 tokens | 102.3 tok/s]*
+
+---
+
+### Prompt: "The history of the Roman Empire begins with"
+
+> The history of the Roman Empire begins with la aquí losña dear no del, tues con suá uno sus which esó inidad está'samos atíair contrados compart el’ cuenta un isos aaciónÍia are afect en experiencia of our. una%  ofadas que usando también toato y te ha sonido historyor that máso history andida theyán “ periodió hasta from .
+
+*[1.95s | 200 tokens | 102.5 tok/s]*
+
+---
+
+## Notes
+
+- AR mode produces semi-coherent English at ~63 tok/s. Topic drift is expected at 20% training completion.
+- SAT mode achieves ~102 tok/s (1.6x faster) but the SAT head has not yet converged.
+- Model: 24 layers, d=1024, 16 heads, rank=128, expansion ratio 2.0x, ALiBi positional encoding.
+- Training target: 35.7B tokens (Chinchilla-optimal for 698M params).