cadif-lever-llama3.1-8b-noinst

A LoRA adapter for Llama-3.1-8B, trained as the lesion arm: instruction blanked from the training query of the interventional arm of the CADIF study on instruction-following retrieval.

What this is

The matched twin of instq: same rows, negatives, steps, schedule and seed, with the instruction removed from the training query. The dataset ships instructed and bare queries as separate columns over identical rows, so the lever is a string swap.

Scored through the CADIF assay on FollowIR's full 104-query census (pool 1000), this adapter records a composition contrast of Delta_comp = +0.0387. Delta_comp is the within-model contrast between the real instruction and a placebo that keeps every word and shuffles their order; it is pinned to exactly zero for a bag-of-words ranker by construction, so a positive value is evidence that word order was read rather than vocabulary matched.

Note that this arm is NOT at the lexical floor: +0.0387 has an interval excluding the exact zero, so contrastive training with hard negatives alone already installs some order sensitivity.

Why it exists

The census in the paper is observational. This arm makes it interventional: if a follower can be constructed on demand from a general-purpose base whose own Delta_comp is at the floor, then a null verdict elsewhere is a property of the system rather than an instrument failing to fire.

Training

• Base: NousResearch/Meta-Llama-3.1-8B (ungated mirror of meta-llama/Llama-3.1-8B)
• LoRA r=32, alpha=64, 7 projection targets; LR 1e-4; temperature 0.01
• 128 in-batch negatives gathered across 4 GPUs; train_group_size 16
• Data: samaya-ai/msmarco-w-instructions, has_instruction subset, N=100k frozen at seed 42
• Hardware: 4x H200 (141 GB), DDP, bf16, non-reentrant gradient checkpointing
• Released checkpoint: checkpoint-600

Scope conditions

A CADIF verdict is a property of a model under a prompt template, on a benchmark, at a pool depth, and all three move verdicts. Numbers above are FollowIR, pool 1000, under the template described in the paper. Do not transport the magnitude across pool depths without rescaling.

Licence and attribution

Derived from Meta's Llama-3.1-8B and therefore governed by the Llama 3.1 Community License. The name retains "llama3.1" as that licence requires for derivative models. Built with Llama.

Citation

See the CADIF paper and the released per-query arrays (tardellirs/cadif-followir-results, DOI 10.57967/hf/9547) and code (DOI 10.5281/zenodo.21297047).