ZhiXia HTLM Action v2 — LFM2.5-350M
This is a merged BF16 Transformers checkpoint for ZhiXia's constrained browser-grounding action contract. It combines the pinned LiquidAI/LFM2.5-350M revision 9e6c6ccf47cd318696e137d381a7ded8fe4df09f with the ZhiXia LoRA v7 adapter trained on 2026-10-05.
Intended use
The model receives one atomic browser instruction and at most 64 compact visible-element candidates. A combobox may expose an options array. The model emits exactly one compact JSON action:
{"type":"click","index":0}
{"type":"type","index":1,"text":"袜子","submit":false}
{"type":"select","index":2,"value":"China"}
For select, the value must copy one exposed option exactly; it must not translate, normalize, or paraphrase the option. This model is a decision component, not an autonomous browser runtime. A trusted host must validate the schema and policy, rebind the selected index to a fresh live element reference, execute the action, and verify the post-action state.
Training
- Contract:
htlm-action-v2 - Data: 2,530 synthetic records with explicit training permission and PII review markers
- Split: 1,754 train / 388 development / 388 held-out test records, isolated by semantic group
- Method: LoRA,
r=16,alpha=32, 3 epochs, 330 update steps - Trainable adapter parameters: 5,996,544
- Training hardware: NVIDIA GeForce RTX 4090
- Frameworks: PyTorch 2.6.0+cu124, Transformers 5.18.0, PEFT 0.21.2
No real browsing traces, credentials, cookies, runtime references, full DOM captures, or approval state were used in this training run.
Evaluation
Held-out seed-v8-select-robust results:
| Metric | Result |
|---|---|
| Schema validity | 100% |
| Target-element exact match | 100% |
| Full-action exact match | 98.97% (384/388) |
| Click full-action exact match | 100% (120/120) |
| Type full-action exact match | 100% (108/108) |
| Select full-action exact match | 97.50% (156/160) |
The unmodified base model scored 3.09% full-action exact match on the same split. This checkpoint also scored 100% on a separate 340-record regression split that included prior Newest, CNY, and arbitrary type-value failures.
Known limitations
- Four held-out failures confused
Price: Low to HighwithPrice: High to Lowfor the Chinese instruction “按最低价格排序”. - Results are from synthetic, offline action evaluation. They do not establish real-site task completion, safety, calibration, latency, or robustness to adversarial page content.
- The model has not passed ZhiXia's real-browser replay, security evaluation, Shadow, Canary, or default-release gates.
- Never let model output bypass host-side authorization, policy checks, live-reference rebinding, or post-action verification.
- Do not use this checkpoint for general knowledge, coding, or unconstrained autonomous actions.
Loading
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "hugoc8/zhixia-htlm-action-v2-lfm2.5-350m"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype="auto", device_map="auto")
Use the exact htlm-action-v2 system prompt and message rendering implemented by the ZhiXia training/runtime contract. Greedy decoding is recommended for action evaluation.
License and attribution
This derivative checkpoint includes and remains subject to the LFM Open License v1.0 supplied in LICENSE. Liquid AI retains ownership of the original LFM model. ZhiXia's modifications consist of the task-specific fine-tuning contribution and supporting metadata. Review the license, including attribution and commercial-use conditions, before use or redistribution.