Qwen2.5-1.5B-Instruct-NLA-L18-av/README.md

license, base_model, library_name, tags

license	base_model	library_name	tags
apache-2.0	Qwen/Qwen2.5-1.5B-Instruct	transformers	interpretability natural-language-autoencoder activation-verbalizer qwen2.5

Qwen2.5-1.5B-Instruct NLA L18 — AV (Activation Verbalizer)

The AV half of a Natural Language Autoencoder trained on the layer-18 residual stream of Qwen/Qwen2.5-1.5B-Instruct. An activation is L2-normalised, scaled by injection_scale, and spliced into the embedding slot of the single injection token in a fixed prompt; the model then generates a free-text description of what the activation encodes.

This repo mirrors the layout of kitft/Llama-3.3-70B-NLA-L53-av: full merged weights at the root (LoRA already folded in — load with plain AutoModelForCausalLM, no PEFT needed) and an nla_meta.yaml sidecar (schema v2) at the root. The AR half lives in the sibling repo dormantx/Qwen2.5-1.5B-Instruct-NLA-L18-ar. The original LoRA-adapter release (both halves in one repo, av/ + ar/ subdirs) is dormantx/Qwen2.5-1.5B-Instruct-NLA-L18; weights here are numerically identical to that release (merged-vs-adapter greedy generations match exactly).

Notes for loaders expecting the 70B layout:

• Qwen2.5-1.5B ties embeddings (tie_word_embeddings: true), so there is no separate lm_head.weight key in the shards; from_pretrained re-ties automatically.
• The AV prompt template in nla_meta.yaml uses the {injection_char} placeholder; the injection token is <|image_pad|> (id 151655), a registered special token that tokenizes atomically in context.
• mse_scale: 1.0 — AR targets in this pipeline are L2-normalised to unit norm, not sqrt(d_model).

Usage

import torch, torch.nn.functional as F, yaml
from huggingface_hub import hf_hub_download
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "dormantx/Qwen2.5-1.5B-Instruct-NLA-L18-av"
meta = yaml.safe_load(open(hf_hub_download(repo, "nla_meta.yaml")))
tok = AutoTokenizer.from_pretrained(repo)
av = AutoModelForCausalLM.from_pretrained(repo, torch_dtype=torch.bfloat16).cuda().eval()

prompt = meta["prompt_templates"]["av"].format(injection_char=meta["tokens"]["injection_char"])
ids = tok(prompt, add_special_tokens=False)["input_ids"]
slot = ids.index(meta["tokens"]["injection_token_id"])
emb = av.get_input_embeddings()(torch.tensor(ids).cuda()[None]).clone()

act = ...  # a raw layer-18 residual-stream activation, i.e. hidden_states[18], shape [1536]
emb[:, slot] = F.normalize(act, dim=-1).to(emb.dtype) * meta["extraction"]["injection_scale"]
out = av.generate(inputs_embeds=emb,
                  attention_mask=torch.ones(emb.shape[:2], device=emb.device),
                  max_new_tokens=32, do_sample=False)
print(tok.decode(out[0], skip_special_tokens=True))

Activations must come from hidden_states[18] of the base model (output of block 18, before the final norm) and be passed raw — the injection step does the normalising and rescaling.

Provenance

Pipeline and honesty-check methodology build on kitft/natural_language_autoencoders, kameshkanna/nla-train, and SolshineCode/nla-gemma-4-e2b.