You can deploy this model and start earning money today!

andyx10/
jacobian-lens-qwen3-8b-chess

Name: jacobian-lens-qwen3-8b-chess
Availability: InStock

Views

No views yet

Jacobian lenses for Qwen3-8B and Qwen3-8B-Chess

file	model	target	source layers	N	size
`lenses/qwen3-8b-base/jlens/pile-10k/lens.pt`	`Qwen/Qwen3-8B`	L34	0–33	100	2.28 GB
`lenses/qwen3-8b-base-chesscorpus/jlens/pile-10k/lens.pt`	`Qwen/Qwen3-8B`	L34	0–33	100	2.28 GB
`lenses/qwen3-8b-chess/jlens/pile-10k/lens.pt`	`ljcnju/Qwen3-8B-Chess`	L34	0–33	100	2.28 GB
`lenses/qwen3-8b-chess-chesscorpus/jlens/pile-10k/lens.pt`	`ljcnju/Qwen3-8B-Chess`	L34	0–33	100	2.28 GB

Corpus

corpus	file	source
generic	`corpus/pile250.json`	`NeelNanda/pile-10k`, documents windowed to 128 tokens
chess	`corpus/chess100.json`	`ljcnju/ChessArena_Training_Dataset` (see below)


N	100 prompts
`max_seq_len` / window	128 tokens (exact — every prompt, both corpora)
`skip_first`	16 (attention-sink positions)
`dim_batch`	16
order	dataset order, no shuffle
dtype (fit / stored)	bfloat16 / float32
`jlens` commit	`581d398613e5602a5af361e1c34d3a92ea82ba8e`

Chess Corpus construction:

Source: ljcnju/ChessArena_Training_Dataset, file standard_chess_top_3_game.jsonl Rows are {model, system, prompt, response}.

Only prompt is used, where a prompt looks like:

Current board position in FEN notation:
r1bq2kr/ppppb1pp/... b - - 1 10
Partial move history in UCI notation:f7g8, g5f7, ...
Legal moves in UCI notation: g8f8, g8f7, ...

Usage

python

import transformers, jlens

hf = transformers.AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B").cuda()
tok = transformers.AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")
model = jlens.from_hf(hf, tok)

lens = jlens.JacobianLens.from_pretrained(
    "andyx10/jacobian-lens-qwen3-8b-chess", filename="lenses/qwen3-8b-base/jlens/pile-10k/lens.pt")

lens_logits, model_logits, _ = lens.apply(
    model, "Fact: The currency used in the country shaped like a boot is",
    positions=[-2])
for layer, logits in sorted(lens_logits.items()):
    print(layer, [tok.decode([t]) for t in logits[0].topk(5).indices])

Attribution

Method and reference implementation: Jacobian lens (jlens) by Anthropic PBC, companion code for Verbalizable Representations Form a Global Workspace in Language Models (arXiv:2607.15495, code), Apache-2.0.
Base model: Qwen/Qwen3-8B.
Chess model: ljcnju/Qwen3-8B-Chess, from ChessArena (arXiv:2509.24239).
Fitting corpus: NeelNanda/pile-10k.

These artifacts are derived from the above; their respective licenses apply.

andyx10/jacobian-lens-qwen3-8b-chess