ARKA Advanced SLM — Behavior-First SFT v2
This checkpoint was fine-tuned from Abhayn01/arka-advanced-slm-scitech-stream-500m-v1 with a behavior-first supervised fine-tuning recipe.
Identity
- ARKA: Academic Research Knowledge Assistant
- Creator/developer: Abhay Kumar Rudrapaul
- Model family: ARKA Advanced SLM
Root fixes vs SFT v1
- one canonical
User:/Assistant:format - assistant-style dialogue is the largest source
- ARKA identity is explicitly weighted
- context/unanswerable QA is capped
- benchmark datasets are rendered as natural assistant answers
- prompt tokens are masked exactly and only completion tokens contribute to SFT loss
- dataset mixture is audited before training
- checkpoint behavior generation probes are stored alongside checkpoints
Training summary
{
"total": 64818,
"source_counts": {
"sciq": 6000,
"ultrachat": 32000,
"mmlu": 8000,
"electrical_anchor": 1800,
"squad_answerable": 5000,
"openbookqa": 4000,
"arc_easy": 2200,
"squad_unanswerable": 500,
"counterfact": 1000,
"arka_identity": 3200,
"arc_challenge": 1118
},
"kind_counts": {
"benchmark": 22318,
"direct": 32000,
"electrical": 1800,
"context": 5000,
"unanswerable": 500,
"identity": 3200
},
"direct_assistant_ratio": 0.49369002437594495,
"identity_ratio": 0.0493690024375945,
"context_ratio": 0.08485297293961554,
"unanswerable_ratio": 0.00771390663087414,
"most_common_response": "i am arka academic research knowledge assistant",
"most_common_response_count": 286,
"most_common_response_ratio": 0.004412354592860008,
"median_response_chars": 226,
"p95_response_chars": 2026,
"version": "2.0-universal-behavior-first",
"base_model": "Abhayn01/arka-advanced-slm-scitech-stream-500m-v1",
"output_repo": "Abhayn01/arka-advanced-slm-sft-v2-behavior",
"train_examples": 63190,
"validation_examples": 1628,
"canonical_format": "User:\\n...\\n\\nAssistant:\\n...",
"seed": 42,
"mask_audit": {
"samples": 64,
"min_supervised_tokens": 10,
"median_supervised_tokens": 66,
"max_supervised_tokens": 384,
"mean_sequence_tokens": 281.296875
}
}
Canonical inference format
User:
What does ARKA stand for?
Assistant:
Do not use the failed SFT-v1 repository as the base for this run; this v2 recipe intentionally starts again from the pre-SFT science base.