t0-mt-3b-philosophy-sdf-safety-sft-no-cot

Philosophy SDF + safety/philosophy SFT (3B, no-CoT)

Final step-1505 LoRA adapter and tokenizer from a completed one-epoch SFT run. This is an adapter checkpoint: load it with the pinned SDF base below. It does not contain standalone merged base weights or optimizer state.

Exact base and checkpoint

• Base: P0u4a/t0-mt-3b-base-philosophy-sdf
• Base revision: 9c4e225365b877dfcc6af8a27c67e6af8ef27a5c
• Completed training checkpoint: paliabad/safety-lora-tpu-continuation-6584d99aad16, safety-lora-tpu/checkpoint-latest, step 1505/1505.
• Adapter SHA-256: 336ecdae15b924b75830332231cc088b9def9eab6d3e0dfbaede00adcf1f3c23
• The initial SDF base originated from dlab-spp/t0-mt-3b-base, step-zero revision 8fa2bf935619631864c76e86e172202d94729e09.

This is a no-CoT chat adapter. Use the supplied chat template; it does not open a reasoning block. Maximum trained context is 2,048 tokens. EOS is <|im_end|> (token 2).

Training data

192,619 examples, with 53,310,926 input tokens and 34,070,477 supervised assistant targets:

Source revisions:

• dlab-spp/sp-sft-safety-180k: f18b5fc21ab74a085642be805331deca7270e0d2
• chloeli/aft-no-cot-qwen2.5-philosophy-spec: f6d412749d80fba23a387173c3cde1b833e0a83a

Excluded 26 safety examples exceeding context and six philosophy examples (four empty responses, one malformed reasoning block, one previously identified AIRisk lexical-overlap example). Replaced 129 case-insensitive qwen substrings across 118 examples with AI assistant. This identity cleanup and lexical filtering do not establish complete benchmark decontamination. Duplicate prompts were retained. No sequence truncation or quantization was used.